SGI Introduces World's Densest Server

Twirlip of the Mists writes "Today SGI announced the Origin 3900 server, the world's densest computer. How dense? How about 16 MIPS R14000A processors and 32 GB of RAM in a 4-rack-unit 'superbrick,' for a grand total of 128 processors and 256 GB of RAM in a single rack. That makes the new machine the densest single-system-image computer in the world; it's even denser than most blade systems. Just for fun, the server also includes a whole bunch of 64-bit, 133 MHz PCI-X slots (from 11 up to hundreds and hundreds, depending on configuration). There's coverage of the announcement on ZDNet, CNET, and InfoWorld, as well as on SGI's own site."

Comments

Just imagine a beowulf ....

[dsb3]

Uh, nevermind.

Re:Just imagine a beowulf ....

[jo42]

Query: What do you call a cluster of slashdot Linux geeks?

Response: The boys that cried "Beowulf!".

Re:Just imagine a beowulf ....

[RageEX]

What is it with beowulf clusters? This is a ccNUMA machine people, breaking it appart into a whole bunch of clusters is asinine, it's one-very-large-system. Though often they *partition* a very-large-system into two or more really-large-systems for redundacy or for a specific task.

Re:Just imagine a beowulf ....

[Anonymous Coward]

Query: What do you call a cluster of slashdot Linux geeks?

Response: GayTradition.com

Re:Just imagine a beowulf ....

[myowntrueself]

nonononono
Take it apart and build a beowulf cluster out of the components?

Don't be daft man!

Imagine, say 64 of these puppies all connected together in a beowulf cluster!

Maybe a whole warehouse of them... say the icecream factory next to Weta Digital...
The 3rd LoTR movie would be out in a *week*

Uh, as long as the production team got their asses into gear. :-P

Re:Just imagine a beowulf ....

[r33per]

Aye: I was thinking that these could be quite useful in the creation of the afore mentioned cluster thingy. You beat me to it...

But with such density

[Anonymous Coward]

comes even more thoughts of Beowulf clusters...

SGI's Gettin' Some

[supergumby]

Good to see a non-Intel compatible platform release something interesting these days. What we need is faster, cheaper hardware that makes sense!

Re:SGI's Gettin' Some

[pixelated77]

I think the last adjective ever applied to SGI is "cheap" :) They'll make Ferrari owners feel like part of the proletariat.

Re:SGI's Gettin' Some

[oliverthered]

Hmmm.... Funny maybe, insitefull no.
Lets look at interesting non-intel compatible platforms.
All of that ASCI stuff that IBM does,
All the SGI clusters out there
The Earth Simulator in Japan.
etc.....
ok so intel is cheep, but it you want harder and faster then you can't afford to have cheep.

Re:SGI's Gettin' Some

[Twirlip of the Mists]

What we need is faster, cheaper hardware that makes sense!

The 128-processor Origin 3900 lists for $2.9 million. There's nothing "cheaper" about this. Faster, yeah; this is one of-- not "the," but one of-- the fastest computers in the world. And it's the densest. But it's nowhere near cheap.

Re:SGI's Gettin' Some

[Sean Johnson]

Just wait for the technology to trickle down. You'll be able to get womething on par with this for $3000 in about, oh say......30 yrs.

Re:SGI's Gettin' Some

[Anonymous Coward]

BZZZZZZZT!

It would be nice to see an Intel compatible platform release something interesting these days.

Re:SGI's Gettin' Some

[Twirlip of the Mists]

There are 128 cpu intel/amd solutions that fit in a single rack. I know of at least 3 companies that produce them and they are cheap.

There are a few blade systems that can squeeze 128 or more processors into a rack, but those are blade systems, not single-system-image compute servers. You can't use a blade server to do the job of an Origin 3900. (Of course, the converse is also true; you wouldn't buy an Origin 3900 to do something you could do with a blade server instead.)

SGI tends to produce exactly what the customer wants. It's just that their customer is more often than not the federal government, or a very large corporation. It's not well-known-- in fact, for a time it was classified-- but SGI designed, manufactured, and sold an entire line of what were basically DSP coprocessor units specifically for Lockheed's satellite division. Called the "tensor processing unit," each one was basically an expansion module for the Origin 2000. SGI built it just like a commercial product, complete with documentation and everything, and manufactured them in large quantities. It's just that you couldn't buy them unless you were Lockheed.

It's only when SGI tries to branch out that they do poorly. I don't know WTF they were thinking when they decided to try selling inexpensive (relative to other SGI products) workstations running NT or Linux. That was just insane. But as SGI strips more and more of that BS away, they get closer and closer to being a sound company again.

Re:SGI's Gettin' Some

[AresTheImpaler]

You'll be able to get womething on par with this for $3000 in about, oh say......30 yrs.

that means no doom 3 for mi in about... 30 yrs. how sad... :(

Re:SGI's Gettin' Some

[bmajik]

They're not out of the woods by any means.

History speaks pretty clearly about what happens to companies that marginalize their business into making 1-offs for infinite-budget DoD contracts and agencies. Eventually, projects get cancelled, line items in budgets get axed, and whole departments are re-orged into something different.

Cray, anyone ? Cray-Research basically went under when the Cray-3 contract was axed. They were counting on that single-machine to keep the afloat. They futzed around with GaAs custom process and never got it qutie working right, and then the cold war ended and with it the justification for subsidizing a maker of 1-off supercomputers.

(Incidentally, the purchase of Cray is what really broke SGI's back. 50% more employees, 2% more market cap, and the O2k/O3k technology came from stanford, not Cray) SGI bought itself into the supercomputing space with the cray acquisition, but their sales reps didn't know what to sell... T3, vector, or Origin. It bled the company pretty badly.

Nobody argues that right now, there are some things for which there simply isn't any other rational choice besides SGI. In the early 90s, that was "anything with video, at all". Look how that market has all but vanished for them.
The problem is the number of markets for which SGI is the only choice is shrinking and will continue to shrink. Only the institutions that need to be 1-3 years ahead of the curve will pay the huge markup for it. The big advantage of the O3k system is, as you ponit out, the single-system image. But this is only really advantageous for lazy programmers, and when you're talking 3m for a machine to do scientific or simulatino work, i suspect a lot of the code running on these is very custom, and NOT done by lazy programmers. So the brilliant thinking SGI has put into the hardware can sometimes be beaten by domain-specific software. Eg, lets say that MOSIX and 10Gig ethernet advances to the point that you can build a 1024p 512 node cluster, where the backbone (10Gb ethernet) is constructed in the same hypercube fabric as the numalink cables, and MOSIX can with software emulate the memory/process/thread migration that O3k is doing now....

then will 2.9m for a machine still seem justified ?... a 512 node wintel cluster is cheaper than 2.9m if the node cost is under about 5500. How many x86 boxes do you know of that cost 5500.. even with 2 procs, a few gb of ram, and 4 or 5 10GB ethernet controllers (so that each node is n-way connected in the same hypercube fabric that O2k and o3k provide)

Re:SGI's Gettin' Some

[Anonymous Coward]

The converse of

"You can't use a blade server to do the job of an Origin 3900."

is.

"You can't do the job of an Origin 3900 using a blade server."

not what you said.

Re:SGI's Gettin' Some

[scotch]

Cray is still alive and well, building super computers. Names change, companies get bought, yada, yada, yoda....

Re:SGI's Gettin' Some

[Twirlip of the Mists]

Cray-Research basically went under when the Cray-3 contract was axed.

Cray has already taken more than $25 million in orders for the X1, a computer that hasn't even been built yet. Cray has had a rough time, but they're doing just fine.

lets say that MOSIX and 10Gig ethernet advances

What if it does? Bandwidth between nodes isn't as big a problem as latency in that case. No matter how fast-- in terms of bits per second-- your network transport is, you're always going to have latencies that are a million times higher than node-to-node latencies inside a NUMA system like the Origin. Seriously, a million times; we're talking milliseconds versus nanoseconds here. Your dismissal of single-system-image designs in favor of cluster designs shows a distinct lack of vision on your part, I'm afraid.

then will 2.9m for a machine still seem justified ?

If you set up the hypothetical situation such that the less-expensive system does everything that the more-expensive system can do, then no, of course the more-expensive system isn't justifable. But that's not reality. SGI can deliver 1,024-processor systems right now. You can call them up and place and order for a 512-processor system right out of their main price list. (Bigger systems are special deals, but the 512-processor configuration has its own part number, just like a workstation or a monitor.)

Two or three years from now, when everything you just described is possible, let's see what SGI has in its price book and revisit the question. I imagine the answer then will be the same as the answer now, just with the facts ratched up a few notches. "Yeah," you'll say, "SGI can deliver 8 kiloprocessors for $3 million, but is it justified? A 2 kilonode wintel cluster is cheaper...."

Re:SGI's Gettin' Some

[demi]

ok so intel is cheep, but it you want harder and faster then you can't afford to have cheep.

You got that right--SGI gives it to you fast and hard.

Re:SGI's Gettin' Some

[Twirlip of the Mists]

You can't just flippy-floppy words around and call the result the converse. That trick only works when the proposition is in the form of an "if-then" statement. The converse of "if x then y" is "if y then x."

This is, of course, aside from the fact that I was obviously using the expression "the converse is also true" in the colloquial sense, not the Boolean. Dork.

Re:SGI's Gettin' Some

[Anonymous Coward]

It is cheaper... Compared to what it's competing against. I think the common error in the responses here on slashdot is the assumption that we are the target market for this beast, which is not *even* close. You would never find this in any mainstream, corporate datacenter. Places like ChevronTexaco might buy just one, odds are they'd just indirectly rent time on one from a third party geophysics company like Veritas (no, not that Veritas, the other one). This is aimed squarely at places that currently have a room or floor for their Cray(s) and other Bigass(tm) vector machines... In other words, the CS department would likely never have a need for it at any school, but the physicists, chemists, and computational biologists will be having a knife fight over who gets one for their department first...

Re:SGI's Gettin' Some

[Twirlip of the Mists]

and other Bigass(tm) vector machines

Heh. When I first scanned this, I read it as "BigBLAS(tm)." Funny thing is, either meaning makes perfect sense.

Re:SGI's Gettin' Some

[Anonymous Coward]

OK, you are getting confused here. Cray Research had nothing to do with the Cray-3/4. That was Cray Computer Corporation (CCC)... which did go under. Different companies, same name. CCC was a spun off by Seimour Cray to pursue his dreams of GaAs computing.

Cray Research did the J90,T3,SV1,etc.. etc.. wihtouh Seimour Cray during the 90's.

BTW, 2.9m is justified, a 512 node Wintel machine can not do what this machine does. I.e. Single image system, and decent NUMA performance. And no MOSIX can not even come close to emulate the O3K. Sorry to burst your bubble, but by all means keep on dreaming....

Re:SGI's Gettin' Some

[Anonymous Coward]

Cheap, compared to making an intel equivilent? possibly.

Re:SGI's Gettin' Some

[protohiro1]

I worked in the Denver SGI sales office when all this was going down. The goverment systems half of our office had to be retrofitted for the meetings. We had to have a special conference room installed that was 100% "bug-proof". Obviously, they didn't let us look inside. But it had an intimidating door with all kind of security features on the door. Of course, none of us new what they were doing, (and I still don't) but I do know that the big customer was Lockheed. All the staff in that branch had to have background checks and now all the SEs and sales reps have a clearance higher than the president (in regards to the stuff this is for). Based on your post, it sounds like they were making some cool stuff.

Re:SGI's Gettin' Some

[jericho4.0]

We might not ever be able to afford something like this. Big iron is not only fast at crunching nubers, but fast at getting that data in and out of the proccessers.

In terms of IO, our desktops can still be beaten hands down by 30 year old systems.

But then again, maybe someone will perfect some super cheap optical system and desktops will be putting out a 1000 times the IO of this thing...

Re:SGI's Gettin' Some

[billd]

ok so intel is cheep, but it you want harder and faster then you can't afford to have cheep.

Cheep cheep! goes the newly hatched chick

Cheap cheap! goes the Hong Kong shopkeeper

There are no degrees of pedanticism.

Re:SGI's Gettin' Some

[Anonymous Coward]

like i care

Re:SGI's Gettin' Some

[fgodfrey]

You are pretty off base here on some points:

The Cray 3 did *not* sink Cray Research. The Cray 3 sunk Cray Supercomputer, which was a different company. Cray Research built an aweful lot of machines after the Cray Supercomputer spinoff. As to what, exactly, sunk Cray to the point of the SGI merger, that is a hard question. Some technical issues coupled with lower defense spending and bad management.

The merger with Cray is *NOT* what broke SGI's back. SGI loved to heap that on the former Cray employees, but that's not the case. Incompetant management, impossible business plans, and lack of a vision on what to do with Cray after the purchase is what has sunk SGI.

Another thing that has sunk SGI, other than incompetant upper management under the Beluzzo administration, is what you bring up in the next paragraph - the number of things that you can do *only* on an SGI is shrinking. SGI failed to recognize this.

As for 10-gig-E and MOSIX: a) 10gig-E doesn't really exist yet in a commodity box b) Even when it does, no commodity PC is going to be able to drive it at full speed c) MOSIX is not yet where Irix is in terms of large scale SSI's and d) in the time it takes for all the first three things to get fixed, SGI is going to have something better (or they'll be out of business).

As for your last paragraph, NO commodity PC presently available has the I/O bandwidth to drive even *ONE* 10-gig-E card at full speed, let alone 4. Also, 10-gig-E cards aren't cheap. Sorry, but you're going to be spending $5500/node even for a PC. Will we be there in 5 years? Yeah, probably, but in 5 years, SGI and Cray will be making even bigger systems.

Oh yeah, one last thing - an x86 can't address as much RAM as this thing can because it's a 32 bit and not 64 bit processor.

Re:SGI's Gettin' Some

[bmajik]

i concede having my history and my names confused. good catch.

re: "my vision"

the NUMAlink is not magic, it is a very specific high performance link with low latency a goal. Eventually, "off the shelf" stuff will be performance competitive, and at a lower cost.

I realize that using hardware not optimized for low latency in place of the FLASH-derivative chips that the router boards use will hurt. I'm referring more to being able to emulate in hosted software the illusion that the NUMA layer creates - that you have a huge pool of local memory, when infact you have a number of discrete pools of physically separate memory, and some slick hardware and software that makes them "seem" local unless you happen to have an app which breaks the implemented algorithms.

My point is that the hardware/software hybrid solution doesn't work magic, its very fast custom hardware with some software that makes it plausible to NOT _have_ to treat what is essentially a cluster like a cluster. There is little inherent advantage to having a single-image OS if you can write a library that your user space HPC apps will use that provides for single-memory and thread/data migration across the nodes in a performant way. And i think cheaper non-custom hardware and intelligent software can bring that layer to traditional clusters.

I'm a huge SGI fan, and i respect the O3k for what it is. (and i respected the O2k and O200 before it). I've read the white papers, even the DASH/FLASH ones from stanford.

Today, the O3k lets many apps scale up to a high cpu count and a large memory size, so that naive implementations can tackle larger problems. In the future, i claim that software and faster interconnect hardware (be it some new ethernet or myrinet or something similar) will let you do the same on commodity hardware.

Re:SGI's Gettin' Some

[Twirlip of the Mists]

In the future, i claim that software and faster interconnect hardware (be it some new ethernet or myrinet or something similar) will let you do the same on commodity hardware.

That's fine. But unless something drastic happens, this "in the future" scenario won't happen for many years-- even the fastest external interconnect is a long way from CrayLink/NUMAlink's specs. Assuming that we're both right, that it will happen eventually but not anytime soon, then by the time your scenario comes to pass, SGI or some other supercomputer manufacturer will have responded to customer needs by making some other large-scale single-system-image implementation that is to this new mystery interconnect as CrayLink was to 100BASE-T in 1995.

It's a circle-of-life thing.

Re:SGI's Gettin' Some

[Anonymous Coward]

Supposedly (I was told by a reliable source, but my memory sucks) SGI used one of their newer workstation mainboards as a primary component for some designs on what i think was the mars mission with the robot on wheels.. too stoned and tired to look it up heh

Re:SGI's Gettin' Some

[Strog]

Didn't you read the article about the SGI system running Linux on 64 Itanium 2 processors the other day??

You could always order a Unisys ES7000 with 32 CPUs. They have 3 models. A Xeon one, an Itanium 2 one and one that can use both.

If you don't think 512 processors and 1TB ram in 4 racks is interesting then I wonder what Intel could do for you.

Most of the applications for this type of platform are completely custom or already written for these types of platforms. There is a reason the high end stuff has gone to Sun, Alpha, IBM and SGI. They have much higher scalibility without the latencies introduced by clusters of smaller Intel boxes and they are extremely matured in these areas. Commodity hardware is great but not always the right tool for the job.

Re:SGI's Gettin' Some

[Anonymous Coward]

or instead, by that time, SGI or some other supercomputer manufacturer(s) will be in Chapter 11.

necessary but not sufficient

[Anonymous Coward]

that's the min system spec for Office 2005! start saving now..

Is it such a good new?

[Anonymous Coward]

I remember eading an article on Slashdot some time ago on how processors were becoming so hot that at the current trend, they would be hotter than nuclear reactors by 2025... Maybe it's time to focus less on "denser and denser, more gigaflops per cube centimeters" and more on lessening heat dissipation...

Re:Is it such a good new?

[pyr0]

"...and more on lessening heat dissipation..."

Correct me if I'm wrong, but wouldn't you want to *increase* heat dissipation?

Re:Is it such a good new?

[lemkebeth]

I think you are thinking of AMD and Intel. Now there is a hot chip for you.

Re:Is it such a good new?

[Anonymous Coward]

You're right. Because of irresponsible companies like SGI, rogue countries like Iraq can legally buy these "fission reaction igniters" and get away with it.

Re:Is it such a good new?

[PaleBoy]

Correct me if I'm wrong, but wouldn't you want to *increase* heat dissipation?

What I think we are really going for here is the decreasing of the lack of more heat dissipation.

Oh wait, is today opposite day?

Re:Is it such a good new?

[Anonymous Coward]

Clusters run colder than single processors for the equivelent speed.
SGI makes clusters.

Re:Is it such a good new?

[podperson]

Um you want to have less heat to dissipate, but dissipate it as efficiently as possible...

So you're both right... or wrong. Whatever.

Re:Is it such a good new?

[Twirlip of the Mists]

I remember eading an article on Slashdot some time ago on how processors were becoming so hot that at the current trend, they would be hotter than nuclear reactors by 2025.

When I got up this morning, it was 59 F outside. Now, just after lunch, it's over 65 F. If this trend continues, it will be hot enough to melt lead outside by next spring!

Beware statistical projections.

Re:Is it such a good new?

[Twirlip of the Mists]

That's true-- or at least they did-- but this particular machine is not a cluster. It's a single computer. You can cluster these with your favorite technology, of course, using Myrinet or gigabit Ethernet or what-have-you. ASCI Blue Mountain at Los Alamos National Labs is a cluster of 6,144 Origin 2000-series processors. I guess it'd be the acme of hyperbole to call that system a megacluster, but the name sounds pretty good.

Re:Is it such a good new?

[RageEX]

SGI makes clusters

Wrong.

yeah, but ..

[Anonymous Coward]

Too bad it runs IRIX!

System Requirements

[Soporific]

Isn't that the system requirement for the up and coming Doom III?

~S

Re:System Requirements

[akula1]

It'll probably be about right by the time Duke Nukem Forever is released.

Re:System Requirements

[Anonymous Coward]

Nope, that's only for the GPU. You'll need a Cray T3E, fully populated with processors and memory in order to bring up the demo screen in Doom.

Re:System Requirements

[drinkypoo]

I don't know, but I suspect by the time DNF comes out this won't be enough horsepower to run a server, let alone the game itself.

Re:

[rmohr02]

How about 16 MIPS R14000A processors and 32 GB of RAM in a 4-rack-unit 'superbrick,' for a grand total of 128 processors and 256 GB of RAM in a single rack.

Damn!

Densest server?

[digitalsushi]

Now where do we find the world's densest admin to run it?

Re:Densest server?

[Anonymous Coward]

I'm pretty sure you're asking the right place for volunteers.

Re:Densest server?

[Anonymous Coward]

Put in a request to management :)

Re:Densest server?

[Anonymous Coward]

Somebody with an MCSE on the planet Jupitor?

Re:Densest server?

[Salden]

My Father used to say I was "dense" when I did something stupid. Do I qualify?

Re:Densest server?

[painehope]

I can give you a few good leads, just let me know what class of dense you'd like :
1) the IT guy who blinks, looks at your GNOME desktop, and goes "Is this some kinda skin?"...then complains when you push him out the door and lock it
2) our "systems engineer" guy who couldn't even sort out which node to unplug , ex. :
me : (name omitted), node16 in rack 5 lost a CPU, go throw a spare in and do a warranty call
(nameless) : dude, no problem
( 5 minutes go by, operators call screaming about a job failing)
( i ping node, it's still dead in the water, do a parallel ping on all nodes, yep, jackass knocked out node16 in rack 4 )
me : (nameless)...what rack did you yank node16 from?
(nameless) : when?
me : ten minutes ago
(steam begins to come from behind ears, reaching for blunt object close at hand)
(nameless) : oh, yeah, that one in rack 4, with the bad memory, right?
me : CPU, rack 5, motherfucker, CPU, rack 5!!!
(nameless) : dude, i did what you told me...
( i just pull hair out, do it myself )
3) or the ever-famous "dense" admin who has consumed so many potato chips that his ass is beginning to swallow the chair he's sitting in...

let me know, we just laid off all of the above...

Re:Densest server?

[Anonymous Coward]

thats too easy..
just check the local msce school.

Re:Densest server?

[SEWilco]

What will be really hard is programming it.
It will be awfully difficult to show it how to do things.
Type very slowly, and use a simple language.
(What can be simpler that a language simply called "C"? Oh, yeah.. "B")

Re:Densest server?

[sharkey]

Now where do we find the world's densest admin to run it?

Go to Monster.com and look for "MCSE".

evil Beowulf

[3Y3]

Every time somebody makes yet another Beowulf cluster joke/reference they make baby IT developer Jesus cry.

Re:evil Beowulf

[Anonymous Coward]

black baby jesus approves!

Re:evil Beowulf

[Anonymous Coward]

Hey, what's wrong with Beowulf clusters anyway..?
think they're kinda cool. Probably missing out on someting here..

SGI

[Izang]

"The US list price for a 128-processor supercomputer with 64GB of memory is $2,937,696." I think that I'll stick with my $400 Octane.

Re:SGI

[gfxguy]

Where can I get a $400 octane?

Re:SGI

[Izang]

Ebay.

I purchased an Octane R10K, 18GB HD, 256MB RAM, granite keyboard, granite mouse, GDM-4011P with SI GFX for $400 from eUnicomp.

It's nothing fancy but fun to play with if you have IRIX.

Re:SGI

[Anonymous Coward]

You buy a $400 handgun and go stealing an octane at the nearest graphic shop.

Re:SGI

[Twirlip of the Mists]

You got ripped off. I wouldn't take a machine with SI graphics if you paid me $10 to haul it away. Ugh.

I guess you can still use the Octane as a space-heater, though. That's a plus.

Re:SGI

[user32.ExitWindowsEx]

I got the same thing. I wanted the CPU for my MIPS assembler class...the graphics weren't that important to me. I liked the 'complete system' nature of the deal - my Athlon's monitor doesn't do Sync-On-Green, so I would have needed a second monitor anyway.
Mabye in a couple of years, I'll stick a VPro V8 board in.

Re:SGI

[the gnat]

That's not really fair; while these machines may be useless for running most modern 3D apps, they're still bitchin' X terminals. The latest version of Irix will run great on those things, and the OS and GUI are so smooth overall that the reduced horsepower isn't really a bother. They aren't for everyone, but they're wonderful machines for people who care more about stability and polish than speed. (Some say the same about Mac OS X, but I've found it has far more annoying hangups than Irix does.)

Re:SGI

[user32.ExitWindowsEx]

Oh, and one more thing, I actually got an Indy first...the poor thing took a week to boot compared to the Octane. It also always was thrashing the disk...even with 128 MB RAM. I sold it at cost and got my Octane from eUnicomp.

Re:SGI

[RageEX]

Look on street corners anywhere on the West Coast. Or look in the junk bin of any big Industrial place in the East, Kodak for example. Or if you really must grab one off eBay.

Re:SGI

[RageEX]

A $300 Octane still rocks-out any current sub-$1000 PC in Pro/ENGINEER.

Re:SGI

[Izang]

It's cool to play with.

I agree, SI gfx are "teh suck", however, the prices for anything better are way out of line. A Maximum Impact card set with a a little TRAM for the old Indigo 2 is still around $900. If you want SE gfx add another $400 dollars to the Octane. For the price of Octane MXE card set I could buy a 4-wheeler and a Geforce 4.

My Octane is used as a personal web server, just for kicks. I wanted to add a video cam but the personal video option is still $1000.

IMO any of those cards shouldn't be worth more than $100. SGI pricing blows my mind. Hell, I can remember when they wanted $1200 for IRIX.

So yeah, I kinda got ripped off, but not because of SI gfx. It's because the SGI market is so unfriendly to casual users. I have 4 SGI workstations; Personal Onyx 4400s (trashed), Octane 10K, Indigo 2 10K and an Indigo. I enjoy the geek factor but I wouldn't recommend them to anybody for anything.

Let me play on a $60K Fuel with some kind of usefull and UPDATED software and maybe I'll change my mind.

Re:SGI

[RageEX]

SI/SSI still rips-it up when it comes to CAD. Load up a 2000+ part assy in Pro/ENGINEER and a couple of windows with smaller parts and drawings and can watch the fastest PC crap out. The Octane just asks for more. MaxIMPACT with TRAM installed is $250 if you know where to go. No one pays the prices listed on reseller pages. Notice also that SGI hasn't sold Indigo2s in many many years, all those prices are from resellers/junk collectors so you can't blame SGI for their high price. As per an extra $400 to upgrade to SE, not if you know where to buy. I can get a complete Octane system with SE for under $400. An entire system with MXE can be had for just under $1500. So the prices you're seeing are way way high an don't reflect what a half-way intelligent person would pay. IRIX is not a product, they do not sell it for any amount of money, it is bundled with the hardware. You can purchase media with a basic right to use/no warranty/no liability/no support for $600. Though you'd be wise to just go to eBay and pay $100 - $150 for a set of media. No such thing as a Personal Onyx, unless you've got some unreleased/strange hardware. Perhaps you have a Personal IRIS (very old). But the R4400 would indicate not. Anyway, the other systems you list are still valueable and quite useful. I use various SGIs for all sorts of things and for the money 2nd-hand systems are a great deal, lots of fun, work increadibly well for certain tasks, and they look really cool. Lastly, there is no $60K Fuel. Fuel starts at $11K and can more than double. A maxxed out Octane2 with 2 CPUs and tons of RAM and every option card you can find including DM6 might push the price to $60K. As per usefull software, anything by Discreet is amazing on an Octane2.

Re:SGI

[Izang]

Sorry, Deskside Onyx not personal. Black with a purple stripe.

Thanks for the corrections.

Re:SGI

[UberLame]

SIs are still great for paint, design, and video editing. My ideal would probably be an Octane with dual SI cards, each with TRAM, a digital video option, and an FC-AL XIO card, and FDDI PCI card.

no different...

[Polo]

I saw a 4-mainboard in a 1u rack-mount server at linuxworld several years ago. That's the same density, nut earlier...

I know Kingstar used to have pretty high-density systems that would mount back-to-back.

Re:no different...

[Twirlip of the Mists]

No, it's not. This is a single-system-image server. The 128-processor rack boots a single kernel. (In fact, you can connect four 128-p racks together to make a 512-p system, and larger systems than that are supported under special contract to SGI. I believe NASA Ames has a 1,024-p.)

The four-processor, 1-unit server you talked about stops there: at four processors. You can't compare that to a system that scales to be 256 times that size.

Re:no different...

[Anonymous Coward]

Yeah, and a whole lot cheaper!

Re:no different...

[Polo]

Whoops, I breezed over that qualification for densest server.

I stand corrected.

Re:no different...

[Anonymous Coward]

<scratch />
<pre>
Huh?
 
Single CPU to single CPU, the R1x000's don't do well against P4's, Athlons, and other commoditized chips.  The memory bandwidth of P4's and Athlons exceeds that of the R1x000, and the latencies are lower on the cheaper chips.  Benchmarks from SpecFPrate and SpecINTrate all tell the same story.  The R1x000 is very long in the tooth, and really needs to be killed.
 
The much vaunted scalability of these systems is also quite questionable.  When was the last time you saw one of the systems tools require this many processors?  What about the application software?  Most of the OpenMP codes max out in the 20-30 CPU parallel range.  OpenMP is shared memory.  MPI (which will run very nicely on a cluster), can get you into the hundreds of CPUs of scalability.
 
The only two significant possible reasons why anyone might really want this, are the memory size, and the IO capability.  You cannot do 1 TB on a commodity system, and there are those who need such systems.  Maybe one or two customers.  The I/O issue is one that many in the cluster camp are struggling with: how do you do IO to a single namespace file system, or a collection of network interfaces in a coordinated manner, leveraging the distributed nature of the system, but still being a system wide resource?
 
Once again, few customers will need that.  
 
There is a niche, and some of SGI's stuff will fit in it.  More power to them.
 
Shared memory is not likely to be the wave of the future, the MTBF for components pretty much guarantees this.  Graceful degredation means you have to fix many things when CPUs/memory/disks fail.  This is a non-trivial problem.
 
Just because you can add processors, doesnt mean the system "scales".  Scaling is one of those fast and furious marketing terms that has an actual meaning attached to it, but you have to wipe away the marketing fluff to find it.

Re:no different...

[Twirlip of the Mists]

I'm not going to get into an argument about processors. Your assertion that the R10000 needs to be retired is unfounded and bogus, and not even worth discussing.

The much vaunted scalability of these systems is also quite questionable.

Okay, you're kind of talking out of your ass now. I've personally seen code slam a 768-processor machine. (It was nothing more complex than an image processing demonstration. SGI's ImageVision software library is hand-coded to parallelize across all available processors. Run ImageVision on a machine with 8 processors and you'll see it run on all 8. Run it on a machine with 768 and you'll see it run on all 768. In the demo, the program did a 5x5 convolution on an image of truly gigantic proportions, 100K by 100K pixels, or something.) Companies like SGI build giant-- or what we would consider giant-- computers because people need them.

MPI (which will run very nicely on a cluster), can get you into the hundreds of CPUs of scalability.

Maybe. But with a single system image, you can scale to tens of thousands of CPUs using nothing more complex than sprocs or pthreads. And you don't get shot in the butt by inter-node bandwidth or latency shortcomings.

Just because you can add processors, doesnt mean the system "scales". Scaling is one of those fast and furious marketing terms that has an actual meaning attached to it, but you have to wipe away the marketing fluff to find it.

I hear what you're saying. You seem to be implying that this system doesn't scale, by your definition. Yet you fail to explain what that definition is, exactly. Indicting the whole concept of single-system scalability isn't going to earn you any points here.

Re:no different...

[jCaT]

Yes, NAS at ames should have the 1024p cluster up by now. When I was interning there three years ago (yikes, was it that long??) they had a 512p cluster up, and it was the first of its kind. There was a port of doom that would run on the LCD status panels on the front of it... let me tell you, there's nothing quite like playing doom on a multimillion dollar hunk of iron.

Ok, so it was damn near impossible.. but amusing nonetheless.

Re:no different...

[Twirlip of the Mists]

Let's be clear. These are not clusters. These are single system images. Clusters are composed of a number of separate computers-- each running its own kernel, with its own address space and its own storage-- that can be used together for certain purposes. The computers you're talking about at NASA are single system images, with one running kernel and one address space. A program running on a 1,024-processor system can use one processor, or some, or all 1,024 at the same time merely by spawning sprocs or pthreads, or by calling library code that spawns sprocs or pthreads.

The chief advantage to the developer of programming for a single system image rather than a cluster is that one doesn't have to use an abstraction tool like PVM or MPI to parallelize one's code. The other big advantage is that single system images are always faster than clusters of identical capacity and processor count.

Now if only we could test it...

[carlmenezes]

with the Slashdot effect, we'd see how good those processors really are :)

Re:Now if only we could test it...

[Anonymous Coward]

You're joking, I know. A slashdotting would just consume their pipe (as in network) before reaching the processors. I'm now done stating the obvious. :)

I can't help but ask: +1 , Patriotic

[Anonymous Coward]

Can you imagine a Beowulf cluster of these babies?

Thanks in advance,
Woot.

Remember, only you have the power to oppose the
Grifter-In-Chief

Cheers!

Makes sense

[sunderland56]

They've got the world's densest CEO, after all.....

World's Densest Server

[bstadil]

Not true.

This record goes to Emmanuel at the little bistro on Rue de Bach just off Blvd. St. Michel in Paris.

Re:World's Densest Server

[Anonymous Coward]

Mod up!

Re:World's Densest Server

[mickwd]

Em-Manuel ?

Don't mind him......he's from Barcelona...

Re:World's Densest Server

[Bob McCown]

Que?

Re:World's Densest Server

[Usquebaugh]

Fire, fire!!!

No it's just a test get back in the kitchen.

Re:World's Densest Server

[JPelorat]

Please.. please try to understand before one of us dies.

Re:World's Densest Server

[protohiro1]

Just for the record, the area around Blvd. St. Michel is also known for the world's densest tourists.

Re:World's Densest Server

[David Mazzotta]

Equally untrue. The world's densest server is Anna Kournikova.

Whatever happened to...

[Cap'n Canuck]

Kinda makes the whole *@home thing passe...

Dense?

[The J Kid]

Stupid servers....getting denser all the time...

({:P for the {:P-impaired)

Quake?

[Anonymous Coward]

I wonder if someone will make a laptop version of this. That would impress the IT guys at my office. AND could heat my house all winter....

yeah...

[2Bits]

now imagine running the heaviest hardware with the lightest of the light OS...

Re:yeah...

[Shinobi]

No thanks. I don't want to see neat hardware be badly crippled by running Linux+GCC on it.

Re:yeah...

[RageEX]

Hey yeah, just like an asshat American muscle car/NASCAR lovin' redneck who pulls a V12 out of a Ferrari 265GTB and slaps in an American V8 for 'real power.'

Re:Heating?

[Twirlip of the Mists]

I meant to mention this in my submission, but it slipped my mind. The R14000A only consumes 17 watts of power. Four of them, plus the Bedrock memory controller chip, plus up to 8 GB of RAM, fit on a board inside a 1 RU clearance. Four of them, plus some nifty backplane hardware, fit into a "superbrick," meaning sixteen processors in 4 RU.

As far as heat loading goes, the "superbrick" is basically one big wind tunnel, with giant fans on the front and ventilation out the back. It pumps a lot of heat into the room, but the temperature in and around the CPUs is really pretty low. I think it peaks around 35 C.

Z....

[BigBir3d]

A beefed-up system with 128 processors and 64MB of memory sells for $2.9 million.

64MB eh? Pretty darn impressive! I have 192MB in this here laptop! I will sell it at a "discount" price of only $2000.00

Re:Z....

[jantheman]

ner ner ne ner ner - I've got a gig in mine.

Re:Z....

[Twirlip of the Mists]

Obviously, that should be 64 gigabytes of RAM, not 64 megs.

Interesting thing about this system will be, rather than the maximum RAM capacity, the minimum RAM required. The original Origin 3000 required some minimal amount of RAM-- 256 or 512 MB or something-- for every four processors. I'm not sure if this new model has the same requirement, but I'd imagine that it does. (It's an architectural thing. Every node board has to have some RAM on it, because that node board may be nominated at boot time to act as the boot master, among other reasons.)

If that's true, then a 128-processor system would require a minimum of either 32 or 64 GB of RAM, depending on whether you can put 256 MB on a node board.

Re:Z....

[RageEX]

If that's true, then a 128-processor system would require a minimum of either 32 or 64 GB of RAM, depending on whether you can put 256 MB on a node board.

You can put up to 32GB on a node board/CX brick, a max of 8 CX bricks per rack makes for 256GB per rack, up to 1TB in a 512cpu system (4 racks). Not sure what the minimum is.

Re:Z....

[katarn]

64MB eh? Pretty darn impressive! I have 192MB in this here laptop! I will sell it at a "discount" price of only $2000.00

Heh, pretty funny, but seriously though, this is just another example of why you should take journalism with a grain of salt. In fact, every article gives a different number for this mock system, the only thing which remains the same is the processor count and the price (sort of)

• ZD net says: 128 processors and 64MB of memory sells for $2.9 million
  
• Cnet says: 128 processors and 64GB of memory sells for $2.9 million
  
• Info World says: 128 processors and 4G bytes of memory costs $2.93 million

I'm sure this sort of thing drives CEOs and celebrities nuts. No matter what you say, by the time it works it's way through the press it comes out mis-quoted and making you look like a dope. Since SGI's web site says the system can take up to 256GB of memeory, it's anyone's guess what the original quote was.

Does it include it's own Fire suppression system?

[jander]

As someone who has worked with blades, my first question is what they do about heat... Sure the CPU's may run a little cooler, but at that density, what keeps it from melting???

SGI ram prices ...exception to prove Moore's Law?

[Anonymous Coward]

Quoting the pinheads at ZD, "The base configuration of the Origin 3900 includes four processors and 512MB of memory. A beefed-up system with 128 processors and 64MB of memory sells for $2.9 million."

I know SGI has always been extremely proud ($$$$) of their RAM, but only 64 *M*bytes of RAM and 128 proc's for $2.9M? Geez, they could at least throw in the 512Mb RAM from the base configuration.

Are we talking Homer dense?

[burgburgburg]

D'oh!

Or Dan Quayle dense?

D'ohe!

Superbrick's layout?

[Puu]

How are the 16 MIPS CPUs arranged in the "superbrick"? Added in increments of four? How much cache? What kind of CPU-to-CPU and CPU-to-RAM buses? Are the CPUs on special modules or PCBs? Something like IBM's Power4 quad-chip octuple-core MCMs with the 32MB cache chips? (And could IBM squeeze those together as densely as in the 3900?) I couldn't find any details at the SGI site. (Understandable as this was just the announcement of the official launch.) Thanks for any insight! Educated guesses do fine.

Re:Superbrick's layout?

[Twirlip of the Mists]

(I'm answering these questions off-the-cuff, so if I mistype any details, sorry.)

If you know what a first-generation C-brick looks like, imagine squeezing that board into a one-rack-unit form factor and stacking four of them together.

Each superbrick includes four boards, spaced one unit apart, with four R14Ks, the Bedrock, and some RAM. The boards are connected with an internal eight-port crossbar router, making the superbrick a self-contained 16-processor unit. Externally, the superbrick connects to the base I/O brick via XIO+; the base I/O brick contains stuff like the system disk and the first 11 PCI-X slots.

I'm not positive how the superbricks are configured. Theoretically, you can partially populate them in one-node increments (meaning 4 CPUs and some RAM), but SGI may or may not sell them that way for manufacturing and QA reasons.

I believe the CPUs come with 8 MB of s-cache each.

The CPU-to-CPU and CPU-to-RAM bandwidths vary depending on the topology you're crossing, but I believe the minimum is 1.6 GB/s unidirectional, or 3.2 GB/s bidirectional. Intra-node bandwidths are somewhat higher, I believe.

No, the CPUs are regular single-core MIPS R14000As. They're tiny chips that don't consume much power, so you can really squeeze 'em in there.

Keep an eye on techpubs.sgi.com, because SGI will be releasing the developer and owner docs for the new system there shortly. (By "shortly" I mean as soon as a few hours or as long as a few weeks, depending on when the docs get released.) You'll find all the technical data you want when those docs go up.

Re:Superbrick's layout?

[Puu]

Thank you very much! All clear now :)

Blade/Origin Comparison

[zmalone]

Commenting on how the new Origin systems are denser then any other single image system, and then comparing them to the current blade fad to make your point is a bit silly. Blades are seperate machines (unless they are Sun, in which case they are the current desktop line), this system is a single machine. I'm not entirely certain about this density claim either, doesn't Sun fit 128 processors in a rack with the Fire 15ks?

Re:Blade/Origin Comparison

[Twirlip of the Mists]

Sun fits 106 processors into a rack. They were previously the record holder. The Origin 3900 is considerably denser than the Sun Fire 15K, both in terms of processor count and PCI-X slot count-- though not at the same time, of course.

I compared the density of SGI's system to blade systems because those are widely considered to be the densest computers in the world, with something like 90 or 100 individual one-processor computers per rack. This system is not only dense in terms of pure processor count that most-- not all, but most-- blade servers, but it's also got all the advantages of a single system image for HPC applications.

Re:Blade/Origin Comparison

[beefguts]

The SunFire 15K can have upto 106 processors in a single box. These Super Bricks can have upto 64 processors in a "Brick" then you'll have to connect to another brick which adds to latency which is the bane of parallel computing. So they may physically be more dense but architecturally they are less efficient.

Re:Blade/Origin Comparison

[Twirlip of the Mists]

Close, but no kewpie doll. A superbrick hold 16 processors (not 64; I think that was a typo on your part), and connects externally via NUMAlink to other superbricks. But, if I remember my numbers right, the maximum memory latency across the longest multi-router NUMAlink hop in a 128-processor Origin 3000-series system is less than the normal processor-to-processor latency in the Sun Fire 15K. NUMAlink is incredibly fast. The ratio of local memory latency to remote memory latency is something 1:1.5, as opposed to about 1:10 in IBM's and Sun's big systems.

Hmmmm

[Chrispy1000000 the 2]

You'd still probably need a beowolf cluster of these things to run the future Quake 3 multiplayer server...

Pointless in most datacenters

[cjsnell]

These servers are pointless in most datacenters. In order to fill one rack with this much horsepower, you would need at least two empty racks next to it to compensate for the power draw and (much) increased cooling needs. I would argue that the target market for this equipment is government labs, research institutes and universities--not usually starved for floor space.

Re:Pointless in most datacenters

[Jobe_br]

Just an FYI - the CNet article (linked above) talks about its possible use on oil rigs - that type of mapping usually takes some horsepower and as usual, anything that is sea-based will be somewhat cramped for space!

Re:Pointless in most datacenters

[Twirlip of the Mists]

Right but wrong. The target market for this system is definitely government and university HPC labs, but those labs are definitely short of floor space. Putting more MIPS per floor tile is an important advancement.

Re:Pointless in most datacenters

[jhines]

The CNet article pointed out that the chip draws 17 watts, similar to a notebook chip. That isn't the whole story, but it helps.

Re:Pointless in most datacenters

[Anonymous Coward]

You're obviously talking out of your ass. The MIPS proc is one of the coolest running and least power hungry processors used in servers today. This is one reason SGI is able to cram so many in a small space--it needs less cooling.

Re:Pointless in most datacenters

[afidel]

At 14W per cpu these things are more dense in terms of energy usage then anything else out there as well. So replace that Sun E15K or rack of Athlons and lower your power and cooling bills (which I believe is about 1/4th the cost of most supercomputers over their lifetime).

Re:Pointless in most datacenters

[leeet]

You can also put them in an airplane. Like I mentioned in a previous post, these machines are extremely decent on the heat side.

Re:Pointless in most datacenters

[leeet]

I bet you guys are using dual P3's 1U servers?
Those can get REALLY hot.

Re:Pointless in most datacenters

[buttahead]

I don't know what data centers you use, but getting extra power into a rack isn't difficult. This can be expensive, but if money is not a factor then power draw isn't a problem.

Cooling, on the other hand, could be a problem. There are ways around this as well, though. Open racks allow cool air to blow fully around and through boxes. Closed cabinets are usually open on the bottom and top to allow cool air from under raised floors to get sucked through the cabinet and out the top. Additionally, I have seen google's setup -- they use huge fans to help air circulation -- and they seem to have gotton around the heat problem.

Keep in mind that the most expense in colocation type datacenters is incurred by rack space. (note: In some cases wan links or bandwidth might cost more.) With that said, I'm not following why this is "pointless in most datacenters".

You can argue that the target market is government or research labs, but the points you mention don't help that argument.

Re:Pointless in most datacenters

[aaarrrgggh]

Actually, the spec sheet indicates that it is 8.9kW per rack (2.2kW for Drive arrays). That is on the high side, but liveable. (6kW is the max for "standard" cooling-- you can accommodate up to 10kW with a high delta-T cooling system. Water cooling comes into play after that.)

The value of shrinking it down is (as you allude to) not a real-estate issue, but more about the computing efficiencies of a denser package.

The HP blades (6U) are about 35kW nameplate per rack, with a real load of about 10-11kW. The energy savings of SGI might actually give it some value in comparison!

Re:Pointless in most datacenters

[Anonymous Coward]

Heh, well, petrochemical executives are just about the dumbest entities on the face of the planet (they don't drown looking up into the rain ... usually ... so they've got a leg up on turkeys); however I don't think even they are stupid enough to put a ~3million USD computer anywhere within 500 yards of oil field workers. ;-) Mud and aersolized petroleum distillates probably don't add much to the performance of a machine like that. And this assumes it doesn't get used by a welder to finance his New Orleans prostitute habit.

There's a reason they stick those guys out on the water away from the rest of humanity...

Re:Pointless in most datacenters

[flaming-opus]

While SGIs are almost exclussively used for High performance / technical computing, and not typical datacenter uses, that is for software reasons, not hardware. The Irix operating system has always been more concerned with scalability and performance than with stability, which did a good job of excluding them from the early web serving and a lot of database work, which was then taken over by IBM, HP, SUN, and more recently by linux.

This rack is not, in fact, going to knock out your cooling / power system worse than a typical rack. 128 processors x 17 watts is 2176watts, plus some extra for routers, etc. This pales to what a lot of data centers are doing with racks of dual P4, linux boxes at up to 6400 watts per rack. That is one of the (few) selling points of the MIPS processor. They don't push up the processor speeds like intel has been doing, but they are nice and cool.

Don't get me wrong, I'm no huge fan of SGI supercomputers, but they are a lot more easily deployed than say a NEC or a CRAY vector machine.

Re:Pointless in most datacenters

[Anonymous Coward]

The Irix operating system has always been more concerned with scalability and performance than with stability

A Cray/Origin2000 with 1024 CPUs ran for about 2.5 years without a single error or system failure. It was powered down for upgrade and redeployment. I know people who throw birthday parties for their Indys, some comming up on 3 years old. Stability is not a problem with IRIX and has never been as long as I've been an SGI user. Dust build-up causes more failures than the IRIX.

Re:Pointless in most datacenters

[Anonymous Coward]

SGIs have been put on oil rigs before. Ever see a 2nd-hand system come in from an oil company? Not pretty, but it still runs. Remember that the major physical design constraint of the Onyx systems were that they had to fit down the hatch of a US nuclear submarine. They go in some extreme places.

Re:Pointless in most datacenters

[katarn]

Not entirely correct. SGI processors put out much less heat and take much less current then most other processors. This is why such a system is possible in the first place. I would no be be surprised if the heat production and power use was in the same range as a typical system. After all, you can only extract so much heat from a given area when you just use fans and moving air, no mater who makes the system. If it had some sort of refrigeration system set up, I might agree with you. But it just uses air and fans like all the rest, and there is only so much heat you can remove in that manner.

Re:Pointless in most datacenters

[Anonymous Coward]

Oil rigs don't do a whole lot of mapping. Sure, you can do some logging as you drill down, and maybe some mapping of the area directly below the rig, but even that isn't going to be real tough. A more likely use would be for oil exploration, i.e. geophysical mapping of magnetic, gravity, radiometric, and other anomalies. And even then, there usually isn't a real good reason to do the processing onl location. Just toss it up a satelite link to a ground station, no reason to be carting racks around. That, and oil exploration tends to be a fairly dirty business, even on ships.

Re:Pointless in most datacenters

[thogard]

I worked for a goverment manager that got an aword because she had the most increased MIPS per square foot. I found a unused sun 690 (new in the box) and we set it up to make sure they hadn't busted it before its warranty expired. Its 216 mips in one rack upped her average which was being held down by an almost one mip ibm 3081 that took 1/4 of the data center.

Re:Pointless in most datacenters

[Anonymous Coward]

Umm. O2ks don't scale to 1024. The biggest was a 512P at NASA Ames. Or do you mean a cluster of O2ks?

Re:Pointless in most datacenters

[RageEX]

Untrue and untrue.

ASCI Blue Mountain is an O2K with 6144 CPUs, it lives at LANL.

See also #52 from www.top500.org:

SGI ORIGIN 2000 250 MHz LANL/ACL USA 1999 2048CPUs.

That's 2048CPUs. SGI only sells pre-configed systems up to 512 CPUs, otherwise you have to call them and go out to lunch before they sell you several thousand processors.

Re:Pointless in most datacenters

[jonbrewer]

In order to fill one rack with this much horsepower, you would need at least two empty racks next to it to compensate for the power draw and (much) increased cooling needs.

Consider some facts:

1. the MIPS chips consume only 17 watts
2. sane datacenter engineers put their air handling gear on the roof.

If a 4U 16-processor SGI could outrun our full-sized Origin 2400 for $300,000, I could see buying one.

We have now the world's densest server...

[StormShaman]

Now we need the world's densest sysadmin to run it!

Fluid Dynamics?

[Shadow Wrought]

From the CNet article:

"P&G also uses the [SGI] system to study fluid dynamics in disposable diapers." And...

"A beefed-up system with 128 processors and 64GB of memory sells for $2.9 million."

From my own delusion mind:

"Look, all we need is another 3mil and we can nip this diaper problem in the butt!"

Favorite Quote

[ProtoStar]

From the ZD link:

Procter & Gamble, for example, uses an SGI system to study the aerodynamics of Pringle's potato chips

Re:Favorite Quote

[ProtoStar]

How long until they use a sgi system to determine the radation pattern of a pringles antenna?

Re:Favorite Quote

[kyoko21]

Imagine what the Wright Brothers, Ford, and the makers of Pringles got together back in the early days of flight and found out that all they needed to make man fly was some pringles and an assembly line.

Re:Heating?

[Christopher Thomas]

Wouldnt something that dense have a tendancy to try to burst into flames?

At 17 W/processor, not really. According to one of the many press releases, this is using a 0.13 micron version of one of their older processors clocked at something like 600 MHz. I'd worry about the bus chipset heating up more than the processors.

It's interesting to look at the implications of a design like this. Highly parallel systems tend to be communications-limited, and systems that deal with large workloads tend to be memory-bandwidth-limited in general. All of this points to the processor not being the bottleneck. SGI appears to have designed with this in mind, using processors optimized for power instead of performance to improve density.

Inappropriate!

[DarkHelmet]

With a beowolf cluster of these...

You'd have a core meltdown that's hotter and does more damage than most nuclear weapons.

<ducks>

Re:Inappropriate!

[ProtoStar]

Duck and Cover

Famous Civil Defense film for children in which Bert the Turtle shows what to do in case of atomic attack.

DIVX (29.6 MB)
VCD (92.2 MB)
MPEG2 (244.5 MB)

I'm not sure...

[powerlinekid]

Today SGI announced the Origin 3900 server, the world's densest computer.
Not if it does't run a Microsoft server product.
*ducks*

17 watts!

[simpl3x]

each processor consumes a reasonable amount of elecricity. why have these never been used in anything other than sgi boxen, and cobalt raqs?... neat processors along with arms of course. too bad the world is stuck in the "my processor is faster than thou" mind set. i had thought some years ago that apple would have been well off buying sgi since they have similar markets at the low end of sgi and at the high end of apple.

Re:17 watts!

[binaryDigit]

why have these never been used in anything other than sgi boxen, and cobalt raqs?

Well Playstations (I & II) use them and they are used in many embedded systems because of the low power/decent performance characteristics. On the desktop, they were in a line of NT based workstations put out by NEC, Sony, as well as MIPS itself.

Density by flops?

[LoudMusic]

How about we calculate density by flops or something else useful. I mean, how difficult would it be to cram a butt load of Pentiums in a rack? Yeah well how much calculation can they do?

Lets cruise on over to the Top 500 and use their handy dandy html list to view 'most powerful chip'. This unfortunately requires a little calc work because they failed to include this number in their table.

#1 NEC Earth-Simulator 35,860.00 GFlops using 5,120 Processors -- WOW!

But that's only 7 GFlops per processor ... that thing is mamoth with 5,120 processors.

Now lets look at a little different design ...

#14 Hitachi SR8000-F1/168 1,653.00 GFlops using 168 Processors -- Hot DAMN!!

This is more like it. They're pulling 9.84 GFlops per processor. With their architecture they could pull off the Earth-Simulator's GFlop rate with 3,645 processors - That's 28% less computer doing the same amount of work. Which means if the Earth-Simulator had been constructed with Hitachi's hardware, they could have been pulling 50,380 GFlops in the same cubic footage.

Now this is all rambling that assumes that the processors are similar in size. Which probably isn't true. But they are also getting more power out of less hardware, and it is rare that THAT isn't a bonus. ... ramble ramble ...

Re:Density by flops?

[GigsVT]

But then that's meaningless too if you don't look at cost. If that 28% reduction in CPU count made the cluster cost 5 times more, then it doesn't look so hot unless your space is really worth that much to you, which is true in some cases and not others.

It's really difficult to look at a single metric for these things, because different things are important to different people.

Re:Density by flops?

[LoudMusic]

Yeah, absolutely. That's the argument I always have with Apple's computers. Who cares if they're smaller and faster if they cost twice as much and are obsolete twice as fast.

Good point.

Re:Density by flops?

[yoink!]

Again though you're own evalutation falls a little short. You have little information on how much power each processor requires (in either case, NEC and Hitachi) therefore effeciency is only measurable in the context of what effecient means to the one doing the measuring. Perhaps the Hitachi chips require 10 times more power than the NEC chips... I mean hopefuly NEC wouldn't design an environmental analysis machine that is, itself, a drain on energy sources. Well it would be rather funny.

Re:Density by flops?

[anzha]

Note though: the SGI press release states that it only scales to 512 processors. it looks like they are having problems scaling beyond that. It is probably having to do with the interconnect and SSI approaches they are taking (at a guess).

That means that you will see a peak of around 5 teraflops. The density is impressive for that speed. The peak performance and scalability is not. Speaking from the Supercomputing world that is. It is something to be proud of (for SGI), but if they want to take the SC world by storm they need to scale higher. The high end of machines that will be ordered over the next 5 years are going to be in the 100+ teraflop range for peak performance. (re: Blue Planet)

While most of the market does not care about the very high end systems - they can't afford them - they ARE excellent PR. Bragging rights can go a long way.

Re:Density by flops?

[LoudMusic]

While most of the market does not care about the very high end systems - they can't afford them - they ARE excellent PR. Bragging rights can go a long way.

This is so true, and it's really unfortunate. Consumers base too much of their purchases on how impressive something appears, or how they can use it to impress their clients. This is all in light of actual performance of the item in question.

A good example is LCDs instead of CRTs on the print servers at my office. They're in public areas and they "look cool". Unfortunately, they get about 10 minutes of actual use during the day. What a waste of money.

Re:Density by flops?

[Panaflex]

PC Load Letter

Re:Density by flops?

[moosesocks]

You've also got to remember, the Earth Simulator was COMPLETED in early 2000. This beast was only released today. In addition, a supercomputer requires a good deal of time to design and build.

Think about it. 60 years ago, we could have built a unit this size with less power than a pocket calculator. Why didn't they use Hitachi's 9.84 GFlop processors....

Technology advances over time. Today that's more the case than ever (even with the economic slump). Remember Moore's law.

Of course, they could just buy a HAL9000

Re:Density by flops?

[Anonymous Coward]

Apple? Smaller and faster?

Like what??

Nice integration maybe, but the specs are not that different from any other PC.

Re:Density by flops?

[RageEX]

This is just ignorant.

Re:Density by flops?

[RageEX]

Note though: the SGI press release states that it only scales to 512 processors. it looks like they are having problems scaling beyond that. It is probably having to do with the interconnect and SSI approaches they are taking (at a guess).

No, no, and no. 512 cpus (4 racks) is the bigest a prepackaged system gets. Above that you have to call them to build you a system. SGI built a 6144 cpu system for LANL in 1998. I think it is safe to say that SGI can go as big as your budget will allow and then some, you just need to call them.

MIPS == Low Consumption

[Puu]

It's only 17W per chip, man. Intel/AMD 64-biters are nearly (if not over) 100W, IBM's Power4 is 115W @ 1GHz (IIRC). MIPS/SGI is actually showing the way in lower power number crunchers!

Ding.

[Trespass]

You are correct, sir.

And as an added bonus...

[stand]

...it'll cook a hot dog faster than your microwave oven.

Re:Heating?

[Anonymous Coward]

It must take a lot to cool it, which would make it pretty loud.

Nope, if you bothered to check the spec, you'd see that there are large intake vents in the front, and a 12" window fan in the back operating at a few hundred RPM. That's only about 33dB.

Re:Does it include it's own Fire suppression syste

[Anonymous Coward]

Not a little cooler, but a lot cooler. Intels draw something like 80 watts. IIRC, This draws like a 486.

No

[Vaulter]

This is the current generation hardware. Doom II will require _next generation_ hardware.

Re:No

[EvilBuu]

Christ, if Doom II will require next gen hardware, then we're pretty much screwed as far as Doom III is concerned.

According to the SGI page

[SexyKellyOsbourne]

The Origin 3900 is cooled with a Bryant Air Conditioning coil and condenser system.

Re:According to the SGI page

[Anonymous Coward]

I enjoy how you change words in quotes.

Re:According to the SGI page

[Anonymous Coward]

I enjoy how this ass is still trolling though. I do get a kick out of reading her journal.

too bad the cooling required is as big as a house.

[Brigadier]

I would be afraid to see how much heat those things generate.

Even more density

[citabjockey]

Check out:

http://www.skycomputers.com/hardware/bolt9u.html

You can build a 16 board system in a 9U chasis that has 16x16 or 256 altivecs. With enough of a hurricane of air you could stuff two of these boxes into a 19" inch rack for 512 processors. How about that?

Re:Even more density

[Twirlip of the Mists]

Doesn't count. It's not a single system image. The Origin 3900 is. But apart from that, that's a lot of processors in a small space.

true to SGI style.

[_ph1ux_]

these things still look cool. I have always liked how SGI boxes look....

I would love to see an SGI server case designed by HR Geiger.

Re:true to SGI style.

[Puu]

I would love to see these cases sold empty at furniture shops. I would have one at home, maybe put a refrigerator in it or something...

Re:true to SGI style.

[binaryDigit]

I've seen a couple of old Onyx servers converted into Coke machines.

Re:true to SGI style.

[Twirlip of the Mists]

The Silicon Graphics Refrigerator Project (or: How To Turn a $175.000 High-End SGI Challenge DM Server into a Fridge)

check out this

[Anonymous Coward]

http://www.paradyn.org/cc.html

Amazing....

[leeet]

It's amazing that a company that is trying to survive can acomplish such an amazing breakthrough. SGI is on the edge, yet it can push their technology far beyond the competition.

I wonder what SGI could do if it had the same number of employees Sun or IBM has.

I think that, once again, they prove that they can provide the community with cool and kick ass products.

Congrats SGI, this is just amazing... Other companies should follow.

Re:Amazing....

[Usquebaugh]

WTF...

Let's see who buys this and what is performs like when installed. I like the idea of single image OS.

I doubt however it's going to make a dent in the Super Computer top 100. It sure as hell ain't going to beat that Japanese monster.

The Japanese made an order of magnitude increase in processing power and you think this toy from SGI is leading edge? LOL

SGI is rapidly becoming the transmeta of super computer manafacturers. There product fills a very small niche, yet all the stupid kids like you think they're so neat.

As an aside the open critical component in a supercomputer is memory, fast memory, whith out that it matters not a jot how quickly your processors work. So what is the memory bandwidth of this baby?

Re:Amazing....

[the gnat]

So what is the memory bandwidth of this baby?

12.8GB/s. What, you couldn't read the fucking product info before posting?

Re:Amazing....

[Twirlip of the Mists]

So what is the memory bandwidth of this baby?

Aggregate, 12.8 GB/s. Actual STREAM TRIAD performance will be considerably higher than that. A 64-processor prototype system using this same architecture and Itanium2 CPUs scored the world record STREAM TRIAD benchmark back in early September. There's little argument that the Origin 3000 architecture is among the fastest architectures in the world.

Re:Amazing....

[Usquebaugh]

Rhetorical question.

Go and look up the memory bandwith of the top 100 and see how it compares.

Maybe I cannot read but you cannot reason!

Re:Amazing....

[sql*kitten]

The Japanese made an order of magnitude increase in processing power and you think this toy from SGI is leading edge? LOL

SGI will happily sell you a 512-processor machine if you want one. The innovation in the 3900 is compute power/m^3, not raw power. It just so happens that the Origin 3000 has got the raw power too.

SGI is rapidly becoming the transmeta of super computer manafacturers. There product fills a very small niche, yet all the stupid kids like you think they're so neat.

Don;t write them off so quickly. There are plenty of things that only an SGI can do. There are circa-1993 Indigo2's still on people's desks (being used for things like Gladiator), because even a 2002 PC can't do some of the things they can do. SGI are a niche vendor, true... but so is Mercedes.

As an aside the open critical component in a supercomputer is memory, fast memory, whith out that it matters not a jot how quickly your processors work. So what is the memory bandwidth of this baby?

Put it this way: internal bandwidth in an SGI workstation is 3.2Gb/s. Can your peecee do that?

Re:Amazing....

[valdis]

"I doubt however it's going to make a dent in the Super Computer top 100"

Actually, it probably will. Looking at the Top500 before the announcement, SGI already had several boxes in the top 100. The vendors they're looking to displace are, by and large, HP and IBM, which together comprise most of the top 100 (only 36 of the top 100 aren't HP or IBM boxen). NEC only has 7 in the top 100, and 5 of those are the SX series boxes, a different series than the Earth Simulator system, which appears to have been a one-off box (as opposed to the IBM gear, which are all SP boxes, or the HP systems, which are AlphaServers from the Dec/Compaq line).

"As an aside the open critical component in a supercomputer is memory, fast memory, whith out that it matters not a jot how quickly your processors work. So what is the memory bandwidth of this baby?"

SGI cites 12.8GB/sec aggregate memory bandwidth. Per CX brick. So the total bandwidth on a 512-CPU system would be some 32 times that (reduced somewhat by the 4x4x32 hypercube interconnect on the Numaflex). SGI claimed 152 GBytes/sec I/O throughput on the older 3800 systems.

And yes, that's a single-system-image. But you knew that. :)

Re:Amazing....

[Twirlip of the Mists]

No, don't bother doing that comparison. The 12.8 GB/s figure is the aggregate memory bandwidth across the system topology. If you want an idea of how that translates into real-world numbers, find an Origin 3900 and run the STREAM TRIAD memory benchmark. I'm making an educated guess here, but I'd expect to see a result around or above the 100 GB/s mark, based on the results that SGI has published from similar systems to this one.

Re:Amazing....

[bmajik]

As an indigo^2 hi-impact owner (r4.4k, unfortuneately)...

the I^2 is a lame machine unless you've got impact gfx, at which point its gfx subsystem becomes ok. The phobos ethernet cards in these things cripple the poor EISA bus and slam the cpu. Playing midi on the things is choppy..

in any case, i chose to respond because of the 3.2Gb/s figure. when you say "internal bandwidth", what are you referring to ? The main memory bandwidth on any DDR system is actually about that fast these days. The O2 doesn't have the XIO bus of the octane, and its mostly irrelevant because pc, o2, and Octane all use PCI for the majority of the expansion devices anyhow.

For a while the memory bandwidth of pC's was a big limiting factor, but its actually gotten pretty good lately, especially compared to the desktop stuff of the traditional big iron players (sgi, sun, rs6k).

Re:Amazing....

[Anonymous Coward]

I wonder what SGI could do if it had the same number of employees Sun or IBM has.

Probably less than they do now. There are some serious advantages to keeping the team of engineers lean, highly motivated, and highly skilled.

I have yet to be impressed with Sun's engineers since they switched from COTS hardware.

Re:Amazing....

[RageEX]

Sounds like you've never used a Challenge or an Origin system before. As per the top 100, it will take a large customer who asks for large system (5,000CPUs+) to get SGI on the list. SGI doesn't build these things for the record books they build them for customers. Even a smaller Origin3900 system beats several of the top systems in memory bandwidth. > There product fills a very small niche, yet all > the stupid kids like you think they're so neat. My two Indys fill a niche on my desk and I prefer using them to my PC. I'm guessing you're talking out your ass, what SGI systems do you have on your desk? > As an aside the open critical component in a > supercomputer is memory, fast memory, whith out > that it matters not a jot how quickly your > processors work. So what is the memory > bandwidth of this baby? You can read about the latency in John Mashey's NUMAflex paper, though I beleive they've improved on this. As per bandwidth, I think it is 13.2GB point to point. Total system bandwidth scales linearly with each CX brick and supporting router. Remember though that this is dedicated bandwidth through a crossbar not the shared bandwidth of a bus. Either way I believe SGI has the highest sustained-bandwidth systems currently on the market. Their Origin-Itanium2 systems set a worldrecord 120GB/s on STREAM. Maybe some highend 1-offs beat it but essentially SGI's NUMA systems are the Ferraris of the computing industry.

Re:Amazing....

[be-fan]

Put it this way: internal bandwidth in an SGI workstation is 3.2Gb/s. Can your peecee do that?
>>>>>>>>>>
Um, yes?

Re:Amazing....

[Shinobi]

First of all, on x86 PC's, that's theoretical bandwidth, practical bandwidth is lower, much lower. An Octane or even an O2 can handle bandwidth-demands that would choke a x86 PC. Some examples are compositing, or handling large images. Playing 1280x1024*32bit uncompressed and doing post-processing in real-time is just one example I've seen.

Re:Amazing....

[RageEX]

Please people take the time to understand before you let shiat fall out your mouth. That is 3.2GB/s of peak dedicated bandwidth through a point to point high-speed non-blocking crossbar. SGI systems don't use a bus based topology. That means that you can have 3.2GB/s from CPUMEM at the same time you can have 3.2GB/s from I/ODISK at the same time you can have 3.2GB/s from GFXCPU (sustained 2.4GB/s). Each router supports 8 connections at 3.2GB/s (1.6GB/s down 1.6GB/s up). And this is for their workstations, the new Origins bump things up a notch.

When you see that a PC can do X.XGB/s peak that bandwitch is shared by all subsystems attatched to the bus.

The main CPU in an desktop SGI system is matched just to fit this architecture. Whilst your AMD system is sitting idle 65% of the time waiting for data to get where it's going and dealing with bus contention issues and a crappy MSFT OS.

Read through Ian Mapleson's Octane Architecture paper to get a better idea of the tech. that goes into an SGI workstation. Please also note that the Octane is a system from 1997 and no longer sold (only on eBay), the Octane2 is faster, and the Fuel is much faster though only offers 1 CPU.

Re:Amazing....

[be-fan]

That is 3.2GB/s of peak dedicated bandwidth through a point to point high-speed non-blocking crossbar. SGI systems don't use a bus based topology. That means that you can have 3.2GB/s from CPUMEM at the same time you can have 3.2GB/s from I/ODISK at the same time you can have 3.2GB/s from GFXCPU (sustained 2.4GB/s). Each router supports 8 connections at 3.2GB/s (1.6GB/s down 1.6GB/s up). And this is for their workstations, the new Origins bump things up a notch.
>>>>>>>>
Most current PCs are also built on a point to point topology. Desktop level chipsets will get you up to 4.2 GB/sec of main memory bandwidth, though the other point to point links aren't as fast. However, Hypertransport is fully capable of scaling to 6.4 GB/sec, and you can already get 3.2 GB/sec (if you need it) for the I/O links from an Intel E7500 chipset.

When you see that a PC can do X.XGB/s peak that bandwitch is shared by all subsystems attatched to the bus.
>>>>>>>
Um, no. The bandwidth measurements are for the PtP link between the CPU and Northbridge (memory controller). For Hypertransport, the bandwidth measurements are for each individual link.

The main CPU in an desktop SGI system is matched just to fit this architecture. Whilst your AMD system is sitting idle 65% of the time waiting for data to get where it's going and dealing with bus contention issues and a crappy MSFT OS.
>>>>>>>>>
Um, my systems don't run Microsoft OSs, and the OS has little to do with the I/O architecture of the machine. And given that an Athlon uses a point to point link between both the GPU and the CPU, bus contention doesn't exist! Who's spouting shit now?

Besides, all this talk about crossbars and whatnot is stupid.You originally asked if PCs had 3.2 GB/sec of internal bandwidth. They do. I never said a PC's architecture was comparable to an SGI's. My point was that PC architecture is catching up, and is a lot closer than you'd like to believe.

Re:Amazing....

[Anonymous Coward]

I still use an I2 today, built in 1993 and don't forget it! It still feels faster than my PC, it just won't play quake. Still does Pro/E though and anything else I ask it. Heck, even an old R3000 Indigo from 1990 is still able to playback MP3s while using Sylpheed, WordPerfect, Galeon, Photoshop, and a bunch of Xterms. Comparing a 10 year old SGI to the latest PC the SGI still wins in *some* categories.

World's Densest Server

[Anonymous Coward]

That would be the blockhead who waited on me at McDonald's last night and couldn't get my order straight after four tries.

not as dense as mine !

[jacquesm]

www.clustercompute.com

well, on a per mips basis maybe, but then again I could use faster cpu's today.

Re:not as dense as mine !

[djdavetrouble]

Beware of evil popup from above link.

Re:Don't Imagine a Beowulf Cluster

[Anonymous Coward]

It'd make a hell of a wall, though. Build your house out of these and you don't have to worry about the rising costs of heating this winter.

I'd also like to mention that I enjoy Subway, despite their lack of 'piping hot grits' as a menu item, and give a shout out to the mods, my paint-huffing homeys keeping it real out there in Internet land.

Larry McArthur?!?

[scsirob]

Check out the picture of the machine on SGI's web site. When hovering the mouse over the small picture of the cabinet, it shows "Picture of Larry McArthur". He's looking a lot better these days..

HEAT?

[IEforLinux]

Sounds like this unit also would be the central heating unit for the office complex in which it resides.

Delaying the inevitable?

[afidel]

Is this just delaying the death of SGI or signaling a new focus and niche for the company? I loved the Indy stations back in college and the O2's were amazing in their time, but most of the work those systems could do can now be done on comodity hardware, so SGI had to find a new reason to exist. Whether this system is enough to keep the grim reaper away is left to be seen.

Re:Delaying the inevitable?

[Twirlip of the Mists]

I can't believe this got moderated as "insightful." Crap like Indys and O2s is what put SGI in a bad place to begin with. SGI always had fantastic graphics technology and a kick-ass operating system. When they tried to sell low-end workstations-- Indys and O2s running IRIX, and all the stupid stuff with Intel machines running NT and Linux-- their net revenues went into the toilet. SGI's biggest sources of revenue have always been scientific and technical computing customers, the government, and the petrochemical/geological industries. It's when SGI de-focuses to talk about stuff like PCs with fancy cases or video servers or data mining software that they start to lose their way.

This isn't SGI finding a new reason to exist. This is SGI going back to what has always been one of its best reasons to exist. Over time, SGI's technical lead in graphics has diminished, fueled primarily by (believe it or not) home computer games. But even now, nobody can touch SGI for high-performance scalable servers like the 3900.

Re:Delaying the inevitable?

[sql*kitten]

I can't believe this got moderated as "insightful." Crap like Indys and O2s is what put SGI in a bad place to begin with. SGI always had fantastic graphics technology and a kick-ass operating system. When they tried to sell low-end workstations-- Indys and O2s running IRIX, and all the stupid stuff with Intel machines running NT and Linux-- their net revenues went into the toilet.

Not quite true. After all, in 1994 an Indy had better price/performance than a comparable Pentium system... and a Pentium couldn't touch a fully loaded Indy. With better marketing, SGI could have dominated the high end 2D and low end 3D space, driving out Apple and Intergraph, and continued to hold high-end 3D. I agree that NT was a colossal mistake for them, and they aren't recovered from that mistake next.

It's when SGI de-focuses to talk about stuff like PCs with fancy cases or video servers or data mining software that they start to lose their way.

SGI servers are fantastic for large databases, the features that make them great for rendering and number crunching (high memory bandwidth, very fast disk I/O, single system image) can easily be applied to databases. The Origins should be wiping the floor with Sun's Fire range. It's a marketing failure, not a technology failure.

This isn't SGI finding a new reason to exist. This is SGI going back to what has always been one of its best reasons to exist. Over time, SGI's technical lead in graphics has diminished, fueled primarily by (believe it or not) home computer games. But even now, nobody can touch SGI for high-performance scalable servers like the 3900.

It has diminished true, but it still exists. There isn't a PC that can touch the Fuel workstation, for example.

Re:Delaying the inevitable?

[Anonymous Coward]

I agree... BSD is dieing too! so as PHP...

Go play your games with your GeForce and 133t P4 and educate yourself _before_ commenting on... Super computer companies... hear me? Super computer.

Is this good?

[mozumder]

Why couldn't they just put 100 P-4 Laptops w/1-2GB mem each, minus screen & keyboard, and networked with a switch?

Should be a little cheaper, too...

Re:Is this good?

[questionlp]

That would work for a simple cluster, but there are somethings that it can't do

• Run a single kernel image
• One laptop cannot access another laptop's memory pool directly (or even indirectly)
• CardBus for expansion is quite limited (CardBus is equivalent to a 32-bit/33Mhz PCI bus compared to a 64-bit/100 or 133Mhz PCI-X bus)
• Laptop hard drives are really slow compared to 10 or 15K SCSI/Fibre Channel hard drives in a SAN
• Although the laptop may be faster in some things (compared to one or two of the MIPS processors) it still sucks in raw I/O
• A P4 laptop running at full speed will still consume over 50 watts of power
• Gotta wait until laptops get Gigabit Ethernet (I know the Powerbook G4s have them)
• Laptops have FireWire and/or USB 2.0, but that still maxes out at 400Mbit/sec or 480Mbit/sec respectively
• ... and I doubt the laptops will last as long as the SGI server nor comply with NEBS or other "server" requirements :)

etc.

Densest server

[boatboy]

And here I thought the densest server was that blonde waitress at Waffle House...

Re:Heating?

[Twirlip of the Mists]

I'd worry about the bus chipset heating up more than the processors.

It does. The Bedrock chip is both considerably larger and considerably hotter than the R14000A is. (Bedrock is the memory controller, node crossbar, and "bus" arbitrator.)

As to your other comment, SGI got a lot for their money when they bought Cray back in the mid 90's. They took a lot of good Cray technology-- like crossbar-based NUMA system design principles-- and incorporated them into their large server systems. I believe SGI was the first company-- other than Cray itself-- to break the one-hundred CPU barrier on a single system image. (The T3 series was a monster, but I don't recall exactly how many CPUs you could cram into one.)

I think it was Seymour himself who once said, "A supercomputer is a device for turning compute-bound problems into I/O bound problems."

Cooling Fans = Wind Tunnel

[Brigadier]

Anyone see the large image of this thing. It has like 10 6" Wide cooling fans. Walking by this thing will be like walking by a turbine jet engine. I cant' wait for the readers digest " Sucked in to the Origin 3000 how I survived"

http://www.sgi.com/cgi-bin/download.cgi?/newsroo m/ press_releases/2002/november/images/origin3900_1_j pg.zip

Re:Cooling Fans = Wind Tunnel

[EZCheese]

It's a supercomputer and a centralized home heating system all rolled into home.

Re:Cooling Fans = Wind Tunnel

[c.emmertfoster]

Why the hell was that image zipped? These guys can make a supercomputer that's denser than Jesus, and they don't realize that jpegs are already compressed? Makes you wonder about those big-brains at SGI. Wierd.

Yawn . . .

[jhylkema]

Nobody's gonna buy this damn thing for two reasons.

First off, the economy is in the shitter. IT spending is very slow; most companies have adopted a "fix-it-when-it-breaks" mentality. And geeks do not and never have made purchasing decisions - PHBs do.

Secondly, SGI, while a great product, will probably be remembered as a company driven into the ground by a stupid CEO who wanted to turn it into Dell.

Geeks need to learn - the best product doesn't always win. In fact, the best product usually does not win, the best-marketed one does. Anyone who doubts this needs to look no further than /.'s favorite whipping boy, Microsoft.

Not so hot

[leeet]

Having the chance to work with a similar machine, I can tell you that the disk arrays (in general) will generate much more heat that the CPU bricks. The CPU bricks a very well ventilated. Hard disks RAIDS (in general) are not so well ventilated and will generate a lot of heat. Maybe they tolerate higher temperature, I don't know.. But it's good though, it keeps a part of the server room a bit warmer when you get too cold :)

We also have a Linux rack and this will get pretty hot too. We had to move the Linux rack next to the A/C blower. I can't really say about other vendors but SGI is doing a good job at cooling their stuff.

SGI Wishes it has the densest server...

[Blowit]

Sorry SGI to break up your party but your claims are not valid... and never will be.

RLX Technologies has the densest server
24 blades in a 3U rack space. That means 336 CPUs in one full Rack.

Eat your heart out SGI. You think you might have something big... when all your doing is hype and making yourselves look like liars. Or is that the gimp who posted the initial message who got OVERZELOUS and posted this pathetic message to /.

Re:SGI Wishes it has the densest server...

[Shinobi]

The difference? The SGI is a single system image, i.e one single computer, that can be used as a server. With the RLX solution, you need to configure it as a cluster, with all it's inherent troubles, to be able to do roughly the same thing, and still not have shared memory etc.

Re:SGI Wishes it has the densest server...

[Anonymous Coward]

As several ppl have pointed out before, the comparision between this SGI server and a rack full of blade servers is like comparing your 3 pinto's to one corvette. I think thats a good analogy.

Maybe you could be nominated for worlds' densest /. poster : - O

Re:SGI Wishes it has the densest server...

[jkovach]

But when you have to get 12 people from point A to point B one mile away, you can do it a lot faster with the three Pintos (assuming they don't explode) than one Corvette. It all depends on the problem you are trying to solve.

Re:SGI Wishes it has the densest server...

[RageEX]

You're in the gutter swearing at the stars, boy.

Re:SGI Wishes it has the densest server...

[Anonymous Coward]

Disclaimer: I'm an SGI employee but I speak only for myself.

What you failed to note is that this a Single System Image machine, unlike those RLX boxes which are multiple OS images.

In short - you're comparing apples and oranges.

Gilad.

Benchmarking the performance

[Anonymous Coward]

I see no mention of benchmarks, despite comparing their system to the industry leading IBM p690.

This is the same level of performance that has resulted in SGI's ass being handed to them on a plate for the last several years. YAWN.

Today, SGI poo-poos faster clock rate CPUs because they do not have one. They say their Itanium2 based system will be out in early 2003. At that time, we can expect SGI marketing to change their tune.

Few computing problems scale to large numbers of lower performance CPUs efficiently (queuing, resource contention, scheduling issues, etc). For most problems, a smaller number of faster CPUs will get the job done more quickly.

I was hoping for more.

Re:Benchmarking the performance

[Wesley Felter]

It looks like the 3900 is just a repackaging of the 3800, so the benchmarks should be the same.

Get a clue

[halfelven]

Your "128 cpu intel/amd solutions that fit in a single rack" are clusters of computers, not single-image supercomputers. Until you understand the difference, go play with the "imagine a Beowulf of these" crowd. :-P

Hint: that difference is why there's a price tag so high on the supercomputers.

Re:Get a clue

[ScottKin]

An interesting comment that seems to be parroted by many in this thread.

sidenote: I began my "career" in Data Processing as a Computer Operator on CDC 6600 and 7600 Supercomputers (the CDC 6600 was the world's first, true "Supercomputer) at Lawrence Berkeley Lab back in the late '70s/early 80's, so I *do* know a little about "Supercomputers". Originally, the only kind of "Supercomputer" was one where there was only one "system image", where a main CPU would parcel-out portions of executable code to PPU's (Peripheral Processing Units) that were either generic in their operational parameters or specialized (Population Matrix Manipulators, Multipliers, Vector Processing). Consequently, there was only a single-image processing model because the additional processors were only given the instructions necessarry for them to perform the duties for which they were made. This is in sharp contrast to current "Clusering" or "Grid" systems, where multiple system images are run on multiple, network-connected PC-based systems or, with the case of "Blade" systems, PC-on-a-card solutions.

MOSIX (http://www.mosix.org) implementations cause the lines between classic SiS solutions and Clusters/Grids to blur even more, due to the way that it handles node-to-node communications and control of the running processes. Yes, it is still a distributed-node processing scheme, but compiled code nearly runs as if it were running on an SiS solution.

While it's true that "Clusters" or "Grids" are significantly different in design and implementation than single-image Supercomputers, the necessity for SiS solutions is becoming less and less clear. The University of Kentucky showed this as far back as 1999 with KLAT2 (http://aggregate.org/KLAT2), as it's LinPACK results placed it on-par with supercomputers rated ~200th at the Top500 List - at only US$41,000.00...and that was back in 1999. Today, there are many "Clusters" and "Grids" that can directly compete (performace-wise) with SGI, IBM & Cray SiS solutions. In fact, a few groundbreaking bits of Clustering magic have come directly from one of the largest single users of SiS systems - Sandia Labs (http://www.sandia.gov)

The lines will continue to blur...

ScottKin

We have the World's densest users!

[Anonymous Cowturd]

We have the World's densest users!

Now if we could only get the two together....

Doesn't RLX make something more 'dense' than this?

[DanEsparza]

Unless I'm smoking crack, RLX actually makes a chassis (with their bladed server) that is more dense than this: http://www.rlx.com/products/products_chassis.php

From their website: "The RLX System 300ex chassis holds 24 ServerBlades in 3U and supports the new ServerBlade 1200i." -- and it's even based on Linus's Transmeta chipset!

Not sure how Sun's server can top this... somebody help me out here.

CYIaBCo%s...

[Anonymous Coward]

Can you imagine a Beowulf cluster of SGI servers SHOVED UP YOUR ASS?!!!



thank you.

Re:CYIaBCo%s...

[3Y3]

I bet it would hurt like hell trying to pass a beowulf cluster out of you're ass.

opps...I made baby jesus cry again.

Not exactly

[halfelven]

SGI makes single-image supercomputers, not computer clusters. There's a big difference between those two, big enough to justify the higher price on the supercomputers.

Mmmm..

[grub]

The US list price for a 128-processor supercomputer with 64GB of memory is $2,937,696.

Will they accept PayPal?

I wonder.....

[kg4czo]

Will it run M$? If so, the licensing alone would probably cost roughly 3 times the computers price, give or take a few million, not to mention upgrade and update contracts. :-p

I read that headline and.....

[Anonymous Coward]

thought it was a story about a blonde working at Hooters....

flawed logic

[Stalke]

If its the densist server than it has off course it has to be more dense than "most" blade servers. It would have to be denser than "all" blade servers.

Why make this?

[wmelick]

OK, nice engineering demo, but is there really a market for this? Like others have said, space is *not* an issue. Cost is. Reliability is. For the handful of places that can afford these things, they also have to be thinking "how many Linux boxes can I buy for 2.9M?" Or maybe I can forklift upgrade my Linux cluster 4 ot 5 times for that same 2.9M.
Perhaps SGI Federal might be able to sell these to power-hungry customers wiht money like NSA and various national labs...but I think the commercial market is almost nil.

And the major factor is that nobody in their right minds buys SGIs anymore. IRIX is nearly dead unless you have legacy apps that require it.

Re:Why make this?

[Anonymous Coward]

Just because your grandma isn't a customer doesn't mean there isn't a market. This is where SGI lives, it's their home. They know what they're doing in this space, as opposed to the workstation market(s). Most of their sales come from selling servers like this, they're high-end is right on track, their low-end is just in poor shape.

The government market is bigger than the commercial but not by a lot. Read through SGIs newsroom fro chris'sake, their main commercial customers are large Fourtune500 companies. The commerical market isn't nil, it's almost a 1/3 of their business.

Your last statement is totally off. SGI customers are not playing Quake and pirating Office, they're using Maya, Pro/E, Unigraphics, Nastran, Mechanica, and other highend scientific/analysis software. For anyone with serious and demanding work to do SGI is one of the top choices, HP or SUN being the others.

From the article

[greenhide]

A beefed-up system with 128 processors and 64MB of memory sells for $2.9 million.

Imagine how much the version with 128 MB must cost!

Re:From the article

[greenhide]

By the way, it was the ZDNet article.

All of the others correctly say 64GB. :-)

Not really

[halfelven]

Your 1U 4CPU mainboard is just a 4CPU system, period. You can connect many of these in a Beowolf/Mosix cluster, but that's a cluster, not a single-image supercomputer.
There are entire classes of problems which cannot be solved on clusters, because of latency (well, they can, but the speed is not good at all). That's why you need single-image systems.

Not only in SGI

[leeet]

MIPS processors are used in a multitude of products. Not just SGI... I remember seeing MIPS based handhelds some time ago.

In fact, SGI doesn't really "own" MIPS anymore, it's an independent company. You can buy their CPU's and start your own company if you wish :)

Re:Not only in SGI

[Twirlip of the Mists]

I have been told-- not officially, but close to it-- that when SGI divested itself of MIPS, it retained all the IP surrounding the R10000 family, which includes the R10000, R12000, R14000, and R14000A, and the rumored R16000, R18000, and R20000 lines. So while MIPS makes more RISC processors than anybody else in the world, only SGI has the rights to develop and produce the really cool ones.

Re:Not only in SGI

[RageEX]

MIPS is also an open architecutre. You can license the ISA and design your own chips. MIPS has one of the cleanest and most compact/RISC-like ISAs around and it is easily extended. MIPS is often the choice of ppl. needing good, fast, cheap chips for their produts (Nintendo, Sony, and other consumer electronics). While MIPS was spun-off from SGI a while back SGI still retains their own design team and the IP surrounding the R10K, R12K, R14K and any future chips. I beleive NEC and/or Toshiba fabs SGIs chips these days.

That's 64GB RAM, not 64MB; it's a typo

[halfelven]

It's a typo.

Woah

[teslatug]

Nice Rack!

Yeah but this is not the same design

[leeet]

SGI's rack is "seen" as a single computer where as that rack will be seen as a multitude of machines (cluster).

And Linux won't scale very well with that many CPU's. I heard stories where 16 cpu's is the fair scaling limit.

Besides, you can "hooks" those SGI racks and build up to 1024 CPU's and it will still be seen as a single computer. Only 1 kernel will be running, not 512 kernels like on Linux (*if* it works!!)

I dont like it

[slasher guy]

It doesn't look all that good, gets to hot to quickly, and is much to big.
For 2.9 million, its not a very good footstool.

But seriously, that thing's got more ram than I've got hd space. Now imagine a... never mind

Re:I dont like it

[questionlp]

Not really... each processor consumes about 17 watts a piece, so at 128 processors, the total power consumption of the processors alone is around 2.2KW. That's not too bad whe compared to a rack of say 42 Itanium2 processors, each requiring about 110W per processor. In processors alone, such a sistem would require 4.62KW, more than double of what the SGI rack would need.

That also doesn't include fans (less power consumption == less heat), logic controllers, memory, etc.

Flops are not everything

[halfelven]

Sure, if you buy a ton of second-hand peecees and glue them together in a Beowulf, you have lots and lots of flops (= CPU power).

But the flops are not everything. The problem with clusters is the network latency when the nodes talk to each other. That latency is small for your average network application, but immense for a supercomputer trying to make all its CPUs talk together. This is why there are entire classes of problems that cannot be solved properly on clusters (non-parallelizable problems).

As opposed to that, an SGI supercomputer has the inter-CPU latency orders of magnitude lower. Same GFlops per total (same CPU power), but certain problems are solved orders of magnitude faster.

That's the power of latency. ;-)

Re:Flops are not everything

[LoudMusic]

No one ever said anything about putting this together as a "network cluster". It's a big ass super computer.

Re:Flops are not everything

[evilors]

''non-parallelizable problems''

What does that have to do with anything ?

A problem is either parallelizable to a degree or not. Assuming a (Theoretical) shared memory computer with N CPUs, there are still problems which you won't be able to parallelize.

The question of how that problem behaves on a message passing system vs a shared memory system comes afterwards ..

no. read the specs

[halfelven]

The MIPS cpu takes only 17 watts to run. That's 10 times lower than your average Intel or AMD.

Re:no. read the specs

[Twirlip of the Mists]

Oh, don't exaggerate. It's closer to 7 times lower. Well, maybe 8. ;-)

Re:no. read the specs

[Anonymous Coward]

I might have misread, but isn't that 8x128 watts?

Re:no. read the specs

[Anonymous Coward]

Are you suggesting the average Intel or AMD chip takes 170 watts to run? Have you gone mad?

I'm certain ...

[Anonymous Coward]

... I had the world's densest server the other day when I simply tried to order a #3 meal at McDonalds. I just gave them the $0.02 after he put my $10 in the register - pretty much fried his brain trying to do the math.

Dedicated Application Computing

[yoink!]

Check out Nvidia's data centers. Beware... windows media format warning.

Notice how many times the word linux is used...

Re:Dedicated Application Computing

[Anonymous Coward]

Strange, it seems they have some mpg format stuff. I wonder what the _big_ problem is with having this particular movie availiable as an mpg??
See here

That's a misconception

[halfelven]

That's a misconception due to the ridiculous amounts of power usually required by Intel and AMD processors.
The MIPS cpus need only 17 watts or so. That's almost 10x lower than Intel/AMD.

No

[halfelven]

Its CPUs need 10x less power than Intel or AMD. ;-)

not that impressed...

[computer_chacham]

Compare this to the 1U Hammer racks that are coming out http://www.newisys.com/NewisysDataSheet.pdf
for a 42U rack, you have 84 processors, with each processor being about two and a half times faster, SPECint2000 1202 vs. 483 and SPECfp2000 1170 vs. 495, with the Hammers in 32bit mode. Each 1U Hammer rack can contain up to 16 GB of memory, which gives a total of 672 GB of total RAM, compared to the 256GB of the Origin 3900. I also wouldn't be surprised if a 42U rack of Hammers ended up costing more like $300,000 than $3,000,000

Re:not that impressed...

[Lord Crc]

It's been pointed out before but:

CPU: 2 AMD Opteron 64-bit capable x86-64 architecture Processors
Memory: 512MB - 16GB ECC, Registered DDR333 SDRAM

Compare that to 128 processors and 256 GB of RAM.

The difference? The AMD rack contains many independent machines, the SGI rack is one machine.

A nice quote from SGI's overview: The architecture is designed to run complex models frequently, and because the entire memory space is shared, large models fit into memory with no programming model restrictions.

You wont have 672 GB of shared memory in your AMD rack!

Just changing focus

[halfelven]

Nope.
SGI does not do graphics workstations anymore. People saw it dissapearing from that arena, and this is how all these ridiculous "SGI is dying" stories appeared.
Nowadays SGI is doing precisely the type of systems described in this Slashdot article: supercomputers. Of course, that has not so much visibility in the home computers market ;-)

Re:Just changing focus

[Twirlip of the Mists]

Well, in all fairness, SGI is still selling the hell out of Fuel and Octane2. And, unless they've cancelled their plans in the last few months-- a distinct possibility-- they've got something new and interesting coming down the pipe. Code-named Chimera, it's supposed to apply the same principles of scalability and modularity to a single-user workstation that they applied years ago to servers and supercomputers. So SGI's workstation business isn't dead, per se, they're just scaling it back to where it needs to be. I don't think you'll ever see another SGI PC. Thank heavens.

Re:Just changing focus

[RageEX]

For the high end I'd like to see 2U workstations using NUMAflex (same chipset from O3x00), with one or two CPUs, room for 16GB RAM, and better-than-VPRO graphics which can be repurposed and racked up to form a large SiS ccNUMA system with the ability to chain together each units graphics cards in a manner similar to InfinitePerformance graphics as seen on Onyx300/3000. HP sells Itanium2 2U systems simliar to what I've described except that when racked up they form a cluster and not a large NUMA machine, and they have nothing like IP graphics.

For the low end I'd like to see a compactified Fuel or similar using either MIPS or Itanium2 (which ever is cheaper) and V12 graphics for under $6000.

Sweet speculation ...

Re:Just changing focus

[Twirlip of the Mists]

You're closer than you might realize, except for the part about the low end. SGI might have plans for a new low-end workstation in that price range, but if they do I'm not aware of it, and I can't say I'd be happy about it if they did.

But for the new stuff... think about the Origin 300. Just read the tech documents on it, and think about it for a little while.

Because that's a cluster, not a supercomputer

[halfelven]

You compare apples and mudpies. An SGI supercomputer is a single-image system, where all 500...1000 CPUs sit on the same bus (therefore talk extremely fast to each other) and share the same memory.
A cluster is different: the CPUs are effectively like islands, without a good visibility into each other's data, with huge latencies when talking to each other.
Yes, i'm aware of Mosix, but that does not solve the latency problem.

I'll say no

[halfelven]

Don't worry, when the economy is in the shitter, there's always your regular three-letter agency to buy those supercomputers. ;-)

Not the world's densest

[Sivar]

Today SGI announced the Origin 3900 server, the world's densest computer. How dense? How about 16 MIPS R14000A processors and 32 GB of RAM in a 4-rack-unit 'superbrick,'.

The server may be SGI's densest, but at least as far as processing power, it is not the densest. As a counterexample, the above configuration has four processors per unit. Many vendors sell 1U Athlon servers in which each unit holds two dual Athlon systems (four processors per unit), and I can assure you that an AthlonMP 2200+ is quite a bit faster than a MIPS R14000 @ 600MHz.
True, those two Athlon systems aren't a single server, but we're talking density here.
Regardless, SGI does have the Athlon beaten hands down on memory per unit.

Re:Not the world's densest

[BenjyD]

and I can assure you that an AthlonMP 2200+ is quite a bit faster than a MIPS R14000 @ 600MHz.

from Spec.org:
specfp2000 Peak Base:
Advanced Micro Devices Epox 8KHA+ Motherboard, AMD Athlon (TM) XP 2200+ 671 624
SGI SGI Origin 300 1X 600MHz R14000A 495 472

Yep, them Athlons are damn fast. At ~60W per processor, a rack of 128 Athlons would consume 7.7kW for the processors alone, so I guess you'd need some decent air-con.

Re:Not the world's densest

[RageEX]

What you quoted is wrong. 1 CX Brick = 4U = 16 CPUs. That's 4 CPUs per 1U of space. A single rack can take 4 CX Bricks = 128 CPUs in 16U of space. That rack also contains an R brick (router), various I/O bricks, disk arrays, and you can squeeze in several InfiniteReality4 pipes. Then you can take 4 racks and combine them for 512CPUs (4 CPUs per 1U, 128CPUs taking up 16U per rack). All one shared memory system. The Athlon system you mention is packing 2 CPUs per 1U of space (half as dense per U), probably 1 or 2 drives, maybe room for 1 or 2 expansion cards, and a decent amount of RAM, all filling up one rack. Makes a nice cluster. Everyone knows the Athlon is memory starved. It's not faster than MIPS on many real work loads. Anyway, it doesn't compare to what SGI is selling. The comparison is like saying a riced-up turbo Honda will beat a Ferrari ... well maybe on a drag strip but dragging is for monkeys, track 'em and watch the Ferrari lap the Honda, that's if the Honda doesn't blow it's engine first.

Re:Not the world's densest

[Sivar]

The Athlon system you mention is packing 2 CPUs per 1U of space (half as dense per U)

From the article I quoted:
At a depth of only 15", the A120 can be stacked back-to-back so up to 4 processors, with over thirteen Gigaflops (billion instructions per second) of computing power, can fit in a 1U space per server

Ergo,

4 processors per U.

Everyone knows the Athlon is memory starved.
3.5GB per system should be plenty for most tasks. For those requiring more RAM per CPU or a larger contiguous chunk of RAM, yes, MIPS (and other 64-bit processors) are a better choice. However, I was talking strictly about processing power density, in which the Athlon systems (which fit four processors per rack) wins.
Regarding real-world tasks, there are bound to be a certain number of tasks that MIPS chips accel in. Just being 64-bit makes them faster at dealing with huge numbers (>~4.3 billion), but I doubt they outperform Athlons on most real-world tasks. Of course, MIPS can run Irix, which is a kick-ass Unix implementation (particularly for video and other streaming media), so they do have the advantage there.

Anyway, it doesn't compare to what SGI is selling. It does in processing density. I would not claim that a bunch of Athlons is a superior solution for, say, a streaming video server or a render farm. Regarding the render farm, though, it seems many movie companies would disagree, as many SGI render farms are being dumped in favor of Athlon and Xeon clusters. (mostly for price, of course)

The comparison is like saying a riced-up turbo Honda will beat a Ferrari
I would say that my comparison was more like saying "given a parking lot of size X, the total horse power of all engines in the parking lot is greater when filled with tightly packed BMW M5's than with a single gigantic diesel engine whose driveshaft is driven by multiple smaller diesel engines, all of which share the same fuel tank."
Of course, car/computer analogies all suck. :)

Here's what you're missing

[halfelven]

...the single-system image... is only really advantageous for lazy programmers...

There are entire classes of problems which cannot be solved fast enough on clusters, but only on single-image systems. Anything that cannot be made into a parallel algorithm falls into that category.

...let's say that Mosix and 10Gig ethernet advances to the point...

With networked clusters you're always going to have latencies, orders of magnitude higher than with single-image supercomputers.
Sure, perhaps in 10 or 15 years, we're going to have network latencies as small as those of a PCI bus, but i'm not really talking about future that far. Until then, clusters will be slow for certain problems. Deal with it. :-P

Re:Here's what you're missing

[Fulcrum of Evil]

With networked clusters you're always going to have latencies, orders of magnitude higher than with single-image supercomputers.

While your point abour ethernet latency is valid, you should be aware that, for somewhat more money, you can get 2gb throughput and about 7us latency. More info at myri.com.

The gap between supercomputer and desktop is getting narrower each year. Eventually you will buy your computer by the pound.

Re:Here's what you're missing

[virtual_mps]

There are entire classes of problems which cannot be solved fast enough on clusters, but only on single-image systems. Anything that cannot be made into a parallel algorithm falls into that category.

Of course, anything that cannot be made into a parallel algorithm isn't going to use more than one processor out of a 512p machine, single system image or not. That's not a target audience for this machine.

Re:Here's what you're missing

[bmajik]

im saying that it is possible to fake single-image in software just the way O3k fakes it in hardware.

its called "Numa" for a reason - the memory access is non-uniform, and there are perf penalties for thinking that all memory accesses are created equal. If you have an app where the hardware/software implementation of the page migrator is a performance blocker, then you'll need to do some tuning work to bring your app up to speed. And at that ponit, you're smart enough to write an app-specific best-case distributed memory management system (ala a much improved MOSIX, which is why i mentioned it)

The advantage of the O3k is that to many apps, it looks like a single image machine, and it "is", but it doesn't let you always make the same assumptions as a single image machine.

Re:Here's what you're missing

[halfelven]

Sorry, my bad.
Of course non-parallel problems cannot be solved on multi-CPU systems, single-image or not.
What i wanted to say is, only perfect parallel problems can be solved easily on clusters. If there's any need for a signifficant exchange of information between nodes ("almost-parallel" problems) then clusters are useless, you need a single-image supercomputer.
Unfortunately, many difficult problems (weather prediction, etc.) are almost parallel, hence the need for powerful single-image supercomputers, like the ones SGI makes.

By the time Office 2005 comes out...

[zerofoo]

You'll be able to buy the Origin machine (or something like it) for $999.99 at walmart.

-ted

Re:By the time Office 2005 comes out...

[Anonymous Coward]

Price Rollback! $977.43!

16 MIPS?

[roofingfelt]

I dunno, 16 MIPS doesn't seem so fast to me. If I'm gonna give up 4 rack slots, I'd want to see at least 20 or 30 MIPS.

Re:16 MIPS?

[Anonymous Coward]

You're dense. One CX brick can have up to 16 MIPS processors, as in MIPS Inc. the company that makes the processors not 'Millions of Instructions Per Second.' You've never heard of MIPS and their hardware or the fact that SGI designs their own MIPS CPUS? That's 16 CPUs per CX brick or 128CPUs per rack, or 512CPUs in a prepackaged system, with more on tap. Sheesh!

Go read the theory, dude

[halfelven]

That's a cluster, not a single-image supercomputer. Read again the coments to this article on Slashdot, there are many explanations why a cluster, no matter how many CPUs you throw at it, will never be able to solve entire classes of problems fast enough; to do that, you need a single-image computer, like the SGI stuff.

Up for debate

[CaptainFlyingToaster]

I think I met the world's densist server at the Hayward, CA Applebees:
Me: "What kind of vegitarian options to you have?"
Her: "We have chicken fettucine alfredo."

Re:Up for debate

[Anonymous Coward]

She probably hates vegetarians.

SGI is better focused than they were 5 years ago.

[zerofoo]

Last time I went to a "sales pitch lunch" for SGI stuff in PA a few years ago, I realized I didn't need to buy SGI equipment to do the low-end stuff our company was doing (3d rendering, multimedia....etc). SGI was catering to the low-end and couldn't compete there. Intel and MS were more cost competitive. Needless to say, we bought nothing from them.

The high-end scientific guys were drooling over the Origin machines....and were willing to fork out for the processing power. This should have been a clue to the SGI product planners. I think they are better reading their markets now and this type of focus may actually save SGI.

-ted

Because...

[halfelven]

You need higher density if you want to put it on a ship. If you think no one wants that, think again (homeland security and all that). ;-)

rrrr

[ltwally]

can you imagine your fps in counter-strike!? i could finally break the 1024x768 barrier!!!

what?.. like you wouldn't be using it for the ultimate gamez boxen too

Re:rrrr

[maelstrom]

Well, no. Most FPS wouldn't scale linearly on multi-processor machines anyway (nor do they have support for more than one CPU). IIRC, Quake III supports SMP but only gets about 30% better performance with the second processor enabled.

What about the Connection Machine?

[bluethundr]

It seems that massively parallel computing has gone the way of the Dinosaur what with the advent of more powerful CPUs. But I read that Danny Hillis of MIT and Thinking Machines fame had built a supercomputer called the Connection Machine which housed 65,536 procs each of which lived on the same wafer with dynamic ram and were arranged in a 16-dimensional hypercube array. I don't think the old beastie had nearly as much ram as the new SGI (of course, this machine was 80's vintage). But depending on the physical size of the old box, could this have not been the world's densest computer ever?

Grammar

[AntiTuX]

I hate to be a dick, but dense(st/r) isn't a proper word. More Dense, Less Dense, Most Dense.
It just bothers me when people use poor grammar.

Re:Grammar

[Artifex]

I hate to be a dick, but dense(st/r) isn't a proper word.

They're not? Who says?

It just bothers me when people use poor grammar.

It bothers me when people can't contribute to discussions, except to complain about grammar. How can you even bear to read Slashdot, with all the comments written by people who can barely spell, or by people for whom English is not their primary language?

Re:Heating?

[TroubleMagnet]

HP sold a 128 way SPP-2000 back in '97 or that was a single image NUMA based system. We could go higher than that but nobody was able to afford a system larger than that. It also only ran SPP-UX (from Convex), not HP-UX so the market was kind of specialized. I think Seymour also said that "a supercomputer is one that costs more than a million dollars"

No, it's not.

[Anonymous Coward]

The converse of "You can't use a blade server to do the job of an Origin 3900." is "You can't do the job of an Origin 3900 using a blade server."

No. The converse of "You can't use a blade server to do the job of an Origin 3900" is "If you find that you can't get the job done using a substitute for the Origin 3900, then it must be that the substitute you were trying to use was a blade server."

Re:SGI is better focused than they were 5 years ag

[Twirlip of the Mists]

The change you've noticed is due in no small part to a shake-up at the very highest levels of management. Rick Belluzzo, while he was a decent enough guy personally, really had some screwed-up ideas about what the company ought to be doing. He saw SGI's competitors as being Dell and Compaq and, to a certain extent, Apple. In fact, SGI's competitors were, and are, companies like NEC and, in a way, Cray itself. Rick thought too small, and nearly killed the company in the process.

live from sgi

[iterations]

"...we booted it up and the damn thing kept saying something about 42 and not to forget a towel..."

Re:Heating?

[Twirlip of the Mists]

It must take a lot to cool it, which would make it pretty loud.

Yes, it's pretty loud. It's also six feet tall and bright purple. Wanna talk about what it smells like?

Some people have an uncanny way of zeroing in on the most irrelevant aspects of things. ;-)

it's MIPS, so it's not a problem

[e]

SGI's MIPS chips are engineered to generate magnitudes less heat than Intel or AMD chips. Itanium cores throw 130W or so, whereas SGI engineers all of its cores to fall between 15 - 20W. And if there's one thing SGI knows, it's how to engineer a case. I have an Origin 200 sitting at home, and that thing has enormous heat sinks on each of the CPUs along with three industrial sized fans pushing air through.

e;

Yeah but that's Federal

[wmelick]

Exactly my point. Federal sales might be there and perhaps oil companies (makes sense to put them on a oil exploration ship...I didn't think of that before).

But then again, this is IRIX. No one wants to write code for that anymore. Well, maybe if you paid up the wazoo for contractors. Everything in IRIX is yucky compared to other Unices. And hell, they'll have to carry a lot of spare parts on board too....SGI hardware is utter crap when it comes to reliability.

And if "homeland security" is running on IRIX, that's just a scary notch above Windoze. Oh wait, I'm forgetting this is the Federal Government we're talking about. Of course the wrong decision is the one they go with. And we the taxpayers pay for it all. *sigh*

Re:Yeah but that's Federal

[Anonymous Coward]

> But then again, this is IRIX. No one wants to
> write code for that anymore. Well, maybe if you
> paid up the wazoo for contractors. Everything
> in IRIX is yucky compared to other Unices. And
> hell, they'll have to carry a lot of spare
> parts on board too....SGI hardware is utter
> crap when it comes to reliability.

Troll. What is so wacky about IRIX, it is SVR4 with BSD enhancements. Bet you can't name one specific thing. As per the spare boards, never have I encountered a board failure. Drive failures yes, but on a 320VW not an IRIX machine. I've got two Indys from 1995, an Indigo from 1991, and an Indigo2 from 1993, never a problem, *always* up, never had even one failure.

Changing Components?

[Captain Large Face]

I don't know much about hardware, but wouldn't this be a bit of a pain to sort out if one of the components blows?

More to come

[Grax]

Coming soon:
World's Densest Husband: "You look fine dear" "Sure my secretary is prettier but she isn't as bossy as you dear"

World's Densest Politician: "I did not have sexual relations with that woman" "but we have pictures" "Oh, well then I didn't like it and I didn't inhale"

World's Densest Waitress: "would you like your usual?" "yes, that sounds good" "OK and what would that be?"

* All jokes and partially humourous phrases remain the property of their original author from now til kingdom come.

Have you any direct experience?

[halfelven]

SGI hardware is utter crap when it comes to reliability.

Just how much experience exactly do you have with SGI hardware?
I routinely work with tens of SGI machines of many different types, as well as Linux on Intel (Dell hardware), Sun, etc. And if i were to make a relibility comparison, SGI is the most reliable. Far more reliable than the Wintel/Dell crap anyway.

Re:Have you any direct experience?

[Anonymous Coward]

Quite a bit actualy - over 500
of their machines, over the past
10 years.

I've also dealt directly with Sun,
HP, IBM, DEC (before they were HP) and all
the "PC" vendors out there. In order of HISTORICAL HARDWARE reliability, based
on 15 years of production experience
with thousands of systems and now millions
of hours of uptime:

1) HP (PA/HPUX machines, not their PC's)
2) SGI (Again, not their PC's - Mips)
3) Sun (Again not their Intel machines they
made in the 80's - yes they did moron)
(Sun gets points off for ent. 4500 CRAP!)

4) DEC (pre compaq)
5) All "PCs" - Dell, Gateway etc.

Re:Heating?

[YokuYakuYoukai]

wow, thats cooler then my single athlon xp :O

Thank god...

[Andy Muldowney]

"The new super-brick utilizes the same electronics as previous SGI Origin 3000 systems, enabling easy incorporation into existing systems, as well as simple upgrades."

Oh, thank god. For a minute there I was worried I wasn't going to be able to easily upgrade my old SGI Origin 3000. I guess "simple upgrades" is relative.

Not really

[leeet]

I know that Exxon has some machines on their rigs.

Required reading for NUMA/NUMAflex ...

[RageEX]

I see lots of errors and misunderstandings here. Apparently people have a hard time understanding tech. that is not in thier PC.

NUMAflex is the coolest concept in systems architecture today. I'm eager to see some trickle down into lower-end markets.

Read this before you post:

John Mashey's excellent NUMAflex paper.

How many of that config will they sell?

[Anonymous Coward]

I'm not trying to troll, I'm genuinely curious. Above someone points out that config is $2.9 million, which is a hell of a chunk for anything, let along a single computer (multiproc or not). How many of those do they expect to sell?

Many years ago I was in on the purchase of an IBM mainframe unit, which by the time we were done the system, disk farm, etc etc came out to be (hold on) $60 million. In the end IBM actually "donated" a lot of that hardware (read: zero cost, probably just to get it sold into the field) so we only paid about $10 million finally I think. But I wondered as we were getting crates "How many of these have they made? Are we the first one? Do most of these designs exist only on paper until an order actually happens? If someone else now orders the same thing as us will they get a completely different machine due to the lessons learned from building ours?" You know damn well there are people in the OS division who when we called about trouble said "Unit #x, right, that one has this quirk..."

Certainly not complaining about IBM here either, we easily made 100x that back in our accounting business.

Re:How many of that config will they sell?

[RageEX]

These type of machines are their biggest sellers, so much so that I think their worstation line is suffering. There is a sizeable market for machines of this size and larger. You can browser through their newsroom to see who is buying what and how many.

SGI maintains a sizeable inventory of Origin and Onyx parts along with all their other products. They literally get the components off the shelf.

For the most part SGI systems are highly standard. They only vary when the customer has big $$$ (like Lockheed or the NSA) and asks for specific changes.

Your second point is really good, one gets what one pays for.

Seti!

[NetNinja]

Ahh! My kingdom for a larger Seti machine!

Dense Server?

[jhughes]

I dunno, the waiter at lunch was pretty dense....three times for the order.....

SGI is one step ahead

[Cheese Cracker]

This is the hardware requirements for the next generation of MS Windows... This info must have leaked to SGI. SGI leads the way!

Just how dense is it?

[Admiral Burrito]

Client:
GET / HTTP/1.1
Host: densestserver.sgi.com

Server:
Um... What's that?

Client:
Do you not understand HTTP 1.1?

Server:
Of course I do...?

Client:
Well then,
GET / HTTP/1.1
Host: densestserver.sgi.com

Server:
Okay... Would you like that biggie-sized?

Client:
wtf?

Server:
Oh, you want a web page. Okay, I get it now.

Client:
Great. Now send it, please.

Server:
Send what?

Client:
*sigh* Nevermind.

User:
Huh? What does "500 Server Error: Server too dense" mean?

nah, I already met the world's densest server...

[demonbug]

at Krusty's. He dropped my burger in the fryer, then kept trying to fish it out with his bare hand.

Yeahsurefine...

[Anonymous Coward]

but does light _bend_ around it?

Just what I thought, amateurs.

Technical stuff about the SGI Origin family

[halfelven]

NUMAflex Modular Design Approach

Quote:
SGI's "NUMAflex" (TM) modular design approach builds computer families
with unusual scalability and evolvability characteristics. It partitions
CPU, I/O, and other functions into small, 19" rackmount computing "bricks",
then combines them via efficient, high-speed cache-coherent cabled interconnects,
rather than large backplanes.

Yes Much

[wmelick]

700+ of O2's, 300+ of Origin 2000's, 50+ Origin 200's, 35 Octanes, 3 Onyx2's.

I've seen SGI memory die, processors die, video problems, VICE chip problems, bad NICs on the motherboard, bad Gig cards (eg0's). Failure to power up. We have a constant stream of parts being RMA'd.

The problem is SGI costs you an arm and a leg for hardware support contracts....not to mention their stuff is getting slower and slower as time goes on with no real future. It's just not a good choice anymore unless you have very specific needs and deep pockets.

Re:Yes Much

[Anonymous Coward]

I am told that the support contracts for the O3x line are much cheaper than the O2xs and my experience is that they have been more reliable at each point in their life cycle.

Just Imagine....

[Anonymous Coward]

a Beowulf cluster of......sheesh, this joke is getting OLD!

Re:Heating?

[Anonymous Coward]

That will be 16GB per node-board soon, everyone is waiting for the denser memory.

not quite

[halfelven]

O3K is a single-image system. Go back to reading the theory please, or stop posting nonsense.

Your intuition is somewhat correct regarding latency penalties for access to "remote" memory versus "local" memory in a NUMA system. However, while a Sun *Fire has a latency penalty of 10:1 for such cases, an SGI O3K has a 1.5:1 "penalty"; hardly a penalty at all. Most people do not bother at all about latency; in those rare cases when people try to optimise for NUMA "remote" memory access, they do that for reasons of bandwidth, not latency (like, backing up 1TB in one hour :-D dumb example, i know, but it's the only real one that comes to mind).
Please read this article for more info:

NUMAflex Modular Design Approach

Also remember that Mosix has latency penalties many orders of magnitude worse than NUMA. Difficult to overcome that.
Not to mention that many problems are "almost parallel" - that means there's a need for heavy data exchange between nodes all the time (many weather prediction algorithms, etc.); with those, no matter how smart your programmers are, you just can't workaround a bad latency in a typical cheap network cluster (Beowulf, Mosix) - you simply need a true single image NUMA system like an SGI O3K.

I couldn't help myself...karma be damned!

[dreamchaser]

Imagine a Beowulf cluster of these!

Origin was an SGI product, NOT a Cray product

[Performer Guy]

Wrong, the CRAY-Link was an SGI invention, the ccNUMA architecture and crossbar based Origin architecture was designed build and ready to ship BEFORE SGI bought CRAY. Buying CRAY was a highly questionable move. The Origin would have killed a lot of CRAY's business and SGI sold their sparc based product design to SUN anyway. SGI ended up with a purchase that would have been their growth opportunity, they could have TAKEN huge parts of CRAY's business instead they bought it and the overheads that went with it. SGI never properly rationalized the businesses and joined the orgamizations after that, and they put a complete ass of a Cray manager in charge of the entire company. SGI beat CRAY into submission and instead of killing them, rescued them and put some of their failed management in charge. Dumb bastards.

Re:Origin was an SGI product, NOT a Cray product

[Twirlip of the Mists]

Yeah, I was actually corrected on this matter in a private email just a little while ago. The product known to the world as CrayLink originated at SGI, not at Cray at all. My time with SGI started in '96, which was after the acquisition, so I was just repeating what I was told. My bad.

In light of that revelation, I have to agree. It seems like SGI didn't really get much, if anything, out of the Cray acquisition. Except those cool purple tee shirts that said, "My other computer is a Cray." I've still got mine around here someplace....

Thanks for correcting me, Angus.

Densest server has 336 processors per rack...

[raynet]

RLX Technologies has a server based on Transmeta Crusoe chip and it can hold 24 CPUs in 3U space, giving 336 processors per rack (and 336GB of RAM and 27TB of HDD :)

See promo here..

Re:Heating?

[Hektor_Troy]

Hot damn!

I have to give up my firstborn to pay for heating my appartment, so I don't turn on the heat, and it's a nice and warm 4 C in this room. For you fahrenheit people - your fridge is probably warmer.

Set me up with some ear muffs, and I'll move into that server room in the morning!

Hey - if they install water cooling, do you think they'd mind if I hooked it up to a swimming pool or something? It'd be nice with an indoor 35C swimming pool!

Last Post!

[alpg]

vi is [[13~^[[15~^[[15~^[[19~^[[18~^ a
muk[^[[29~^[[34~^[[26~^[[32~^ch better editor than this emacs. I know
I^[[14~'ll get flamed for this but the truth has to be
said. ^[[D^[[D^[[D^[[D ^[[D^[^[[D^[[D^[[B^
exit ^X^C quit :x :wq dang it :w:w:w :x ^C^C^Z^D
-- Jesper Lauridsen from alt.religion.emacs

- this post brought to you by the Automated Last Post Generator...