Slashdot Mirror
R12K Debuts
Mike Holve, the
guy with the cool website a lot of us have read, wrote in with a link
to an SGI Press Release about
the company using the new R12000 (R12K) processor running at 300 MHz in its Octane workstations. The
chip is binary- and pin- compatible with the R10K, and from the press release, appears
to be relativley inexpensive. Hey, I can dream, can't I? If anyone
sees a press release about the actual chip, submit it and I will be sure
to update the article. Update: 02/22 07:34 by J :Still no sign of a press release, but thanks to
mmoore, we have a nice license plate mask.
Realize that those numbers are list price, which nobody ends up paying (almost nobody, anyway).
You get major discounts.
Can't wait to upgrade our servers.
R12000 + Gigabit Ethernet + Samba = Smooken!!!!!
Does my PII running Linux count as a competitive UNIX system ? :)
Hopefully not.
RC5 is a complete waste of time.
Why is everyone getting excited over a chip, which to upgrade costs $4500?
Did you know that you can buy 80 (EIGHTY) Intel Celeron 300A's (at $56 a piece, lowest price on pricewatch) for $4500? So comparing total compute throughput:
(1) R12000 (80) Celeron 300A's
specInt95 17.3 904 (11.3 * 80)
specFP95 27.5 720.8 (9.01 * 80)
Harness the power of $4500 worth of Celerons.. now that's something to drool over.. none of this $4500 to "upgrade" crap.
p.s. for you nitpickers, yeah.. i do know that $4500 is SRP, not street.. but do the math.. even if the street price was HALF, you're still paying WAY too much
Got some comparisons?
you cannot simply scale the SPEC performance by the number of processors. i do not think that 80-way multi-processing is even *available* with the IA-32 architecture. i suppose you would have to build a beowulf cluster. you must also remember that a computer is much more than one processor. building a network that supports 80 celeron processors would be vastly more expensive than $4500.
http://beowulf.gsfc.nasa.gov/
~ spiderus ~
Erm, you seem to be missing a rather major point here... if you bought those 80 sticks of celery, where would you put them? A good Abit bh6 motherboard for each would be $93*80=$7740, plus an extra 50 each for a case, plus another 20 for a network card to beowulf them (and here I'm assuming that they have no graphics cards and you are using a serial terminal) = $17520 for your cluster, not counting the switches you will need... call it at least $20k, and you still wouldn't be as fast in the real world. Because most tasks are really memory bandwidth limited at that level of performance, and while the R12k is optimized quite well for that, the beowulf cluster sadly isn't. Most tasks aren't all that parallelizable you know...
Our local sales puke sez: Taking origin orders in
later March time frame to ship in late April/early May.
Edwin
Why are you all so impressed?! Machines based on this CPU aren't so great, much less an extremely overpriced, underpowered SGI box. Even though it has some great packages that utilize it's graphics power, you get more bang for less buck when investing in a GNU & [Linux|BSD] system..
As far as the CPU, SPECint 20 or something? haha the hype over this CPU typifies the disillusion you're under
I can't imagine actually buying anything for so much from them... haha suckers!
I partly agree. Linux/BSD have the best price/performance. But Linux is existent more on the low end server market. Linux will not even run on the kind of boxes SGI is producing (128 CPU SGI Origin2000 server comes to mind). And SGIs generally compare quite well to Intel Xeon boxes IMHO.
Sorry, don't compare your sh*tty PCs to SGIs. Thats silly.
As for graphics, Linux sucks, period. It has a very silly 3d hardware support (3DFx only) and there aren't that many usable 3D apps.
~ spiderus ~
Where's Linux 64 bit file system?
What's Intel doing with NUMA?
PCI Bus
Ha!
We generate 1GB+ an hour of data at some of our sites. When processing, files commonly grow bigger than Linux's 2GB limit.
Sigh.
If Linux/Intel could hang with the Big boys we would love to use it as $GI cost so darn much.
Linux/Intel just isn't quite there yet, but SGI was when we purchased our O2000 30 months ago (about a billion years in computer evolution).
The 2GB limit only holds true for 32 bit architectures. 64 bit CPUs can address more memory
the R14000 and the R16000 were -not- cancelled. the 'capitan' and 'beast' projects were cancelled.
~ spiderus ~
2GB limit PCI=LOL was refering to is file size limit, we have files up to 7GB, IRIX/XTF doesn't mind.....
- A.P.
--
"One World, One Web, One Program" - Microsoft Promotional Ad
"Remember when the U.S. had a drug problem, and then we declared a War On Drugs, and now you can't buy drugs anymore?"
I have two SGI Indy's, been a big fan of the Indy since it was released and I scrounged together enough cash to get an entry-level system which I've progressively upgraded over the years.
Where's a good place to buy 2nd hand SGI gear these days? I've considered getting an O2, but I figure if there is a cheap resource for Indigo2's I could always just go that route too...
; -- the corruption of government starts with its secrets. a truly free people keep no secrets. --
>With IA64, Intel may finally catch up to the Rxk,
>UltraSparc, and Alpha chips. Perhaps. But we'll
>have to wait for SGI or HP to make a machine that
>isn't bogged down by its own bus architecture.
What are you talking about? I can't think of a single SGI machine that has bus/IO bottleneck issues. Even the new Visual Workstation has a new superwide bus that can easily handle a mere Xenon.
I read the internet for the articles.
Come on, now. Those CPUs just keep on scaling. You can easily have 128 out of the box, 256 if you ask nicely. The government has a box with over 6,000 MIPS CPUs running... How much do you want?!
The math on the MIPS chips will make anything Intel makes look like roadkill, especially the R10/R12K chips.
How many Celerons can you pop into a box? Now, how many MIPS? Intel can't shake a stick at SGI's SMP. Not to mention that you can't compare a Celeron to an R12K to begin with!
I know what you're saying, or trying to - but you're comparing two totally different ends of the design spectrum. Celerons are entry level CPUs. R12Ks are top of the line. One is meant for low-cost home use the other for seriously heavy duty SMP machines crunching hard numbers, which 160 Celerons couldn't keep up with.
Very good news indeed. As if the Octane machines aren't already totally kick-ass, this announcement comes up.
Someday I might trade in my Indigo2 for an Octane, but even used - an entry level Octane is around $18K, and that's not from SGI, so the trade-up credit of "up to" $7,500 means nada to me (unless of course, I bought the Octane from SGI - way too expensive). Besides, I totally love the Indigo2, especially for 3D (it's a High Impact - realtime 3D!) and video. But some day... Hey, we can all dream, right? :)
Thanks Justin!
P.S. Keep an eye out for a new SGI/Linux page I'm working on. Bunches of cool stuff, especially for you SGI and Linux owners out there. (links, info, pre-compiled binaries, etc.)
More opinions and answers are always good.
For what it's worth, there was nothing there as I posted either... We need a realtime Slashdot! :)
I got my equipment and also highly recommend Reputable Systems. Greg Douglas is a great person to deal with and I'm a very happy customer.
Now, once the port of Linux to MIPS CPUs (or perhaps, more specifically - SGI hardware) is done - and is complete - then you might have an argument - but look also at the hardware that SGI has... You're comparing apples to oranges in an absurd comparison to begin with.
P.S. My SGI runs the GIMP a hell of a lot faster than my Intel box running Linux, thanks...
Now will it crunch RC5 keys faster then say an
intel pentium 200?
The key word being 'relatively'. :-)
Daniel
Hurry up and jump on the individualist bandwagon!
--
Care about electronic freedom? Consider donating to the EFF!
I think it does well in the embeded market because the R4x00 is really damm cheap compaired to any other embeded CPU with similiar performance (or at least it was two years ago, when many of todays products were designed). Go look at the NEC(?) VR4600 prices and tell me the 120Mhz i960 is price competatave! Remember the 4600 has lots of stuff integrated into it as well (not as much as the 68F333, but the 68F333 only runs at 25Mhz!, nor as much as the Cyrex MediaGX, but the MGX is way to costly to see in a cell phone! The MGX's integrated parts are also mostly PC desktop/notebook specific)
Um, what are you trying to say here? As far as I know the MIPS code density is roughly the same as the SPARC, somewhat better then the ALPHA, around the same as the i960. It's only worse then the CISCs, and the x86 CISCs are gennerally to costly and slow for most embeded applications, so that leaves the 68k and CPU32, both of which are fairly slow compaired to a R4x00 (or an R3000!), or any other modernish RISC.
So the MIPS sucks because you don't have to hand-code assembly to get good performance? Does that mean the BMW M3 sucks because it gets good 0-60 times with it's automatic transmision?
I've hand coded SPARC assembly, and while I can do better then gcc, I don't beat Sun's compiler. Unless I happen to be using the VIS instructions (only on the UltraSPARC, and not a part of SPARC V9). Does that mean the SPARC sucks too now?
NOTE: the SPARC and MIPS are both quite simple ISAs, but to get good performance on a modern implmentataion of either you need to pay careful attention to the grouping rules, and data dependencies so your SuperSPARC can issue as close to four instructions per cycle as possable, or your Ultra as close to five as possable (sorry, I don't recall the MIPS maximum execution rates, nor the max rates for the Alpha's either). It's actually harder to hand code fast assembly for these beasts then it was back on the old 68020! Or many other assorted CISCs. Then again it isn't my job anymore, so maybe I'm just bad at it now.
Bonus Alpha trivia: the 21264's register rename file is actually two banks of 40 registers each, and if your instruction ends up referencing registers from diffrent banks there is a one cycle penalty. Because these are rename registers they are assigned at runtime, baised on control and dataflow, including effects of mis-prediction. It's basically impossable for a person to figure out how the registers will get renamed, but a really smart compiler could probbably make some good guesses. Not that I have seen a compiler do that mind you, it's just the kind of stuff modern CPUs do that make hand optmising a tripple bitch and a half of a nightmare.
Ok, I forgot about the ARM. It's code density isn't all that great, what with losing 4 bits to make every instruction conditional (which is cool, but not very dense), and a few more for the shifter that can apply to every ALU output. Even using the DEC Thumb extension (16bit coding) the code density isn't awesum because your back to a two register ISA, and far branches & big constants are a bitch, but yes it does have the best code density of any RISC if you use the Thumb extension. It's also about 20% slower that way, is it not?
Sufficently fast for what? The embeded market is very diverse. The folks replacing complex CAMs in washing machines may be just fine with a Z80 (or Z8!), but the folks doing the PlayStation2 want more comp-u-trons. The PalmPilot has a "mere" 16Mhz 68ECxxx in it, and I hear lots of folks that want something faster! The CPU in my thermostat is probbably just fine as a Z80, but the one controling my ABS breaks would be pushing it as a Z80. A 68K may be just fine to control an elevator, but a PostScript laser printer needs a hell of a lot more CPU! A Z80 in my microwave may be overkill, but you need a lot more then that to control a router with a gigabit backplane. I use to do embedded systems (CoinOp video games), beleve me there are embeded systems, and there are embeded systems! I can come up with a bunch more examples where you want more then a 68K, but I think everyone gets the point.
It's good to see that SGI didn't dump the MIPS processor, and there remains that diversity in the market. If everyone abandoned their own processors it would be a bit of a nightmare.
.. the Intel Lada 1.2 or the Intel Lada 1.3....
Heaven help us if we ever end up faced with one (or two) choice(s) of CPU...
I'm out of my tree just now but please feel free to leave a banana.
A little insect friend of mine out in the desert
tells me that the govs' blue mtn. has been using R12k cpus for some time now; for test purposes, of course >:)
350Mhz Ip31 _me_, baby!
A little bird that chased the fly to make
him lunch sez that Sgi's internal sales crapola lists the 300Mhz octane CAD #'s at roughly 30% faster than its previous R10k release (not the 250s) (with Mxe graphics)
fly on the wall...
(note: all fly info violates no known NDA's)
MIPS is royally screwed now. I mean, what is it?
If you sell it as a high performance
"compiler-friendly" architecture, there is Alpha.
They will be hard-pressed to consistently outperfrom Alpha. An embedded architecture? Well,
a lot of people use it. I suppose that's because
they used Patterson/Hennesy books in their
CPU design class at school. IMHO, MIPS is horribly
miscast as an embedded architecture. Because it
is so streamline, the code density is horrible.
This is the MIPS swan song. As far as I know
there is no next generation MIPS CPU in the
pipeline. The 'Beast' project and the next one
after that are dead. Or is this the 'Beast'?
Doesn't look like it. Beast had a target clock of
around 666Mhz. SGI is probably moving everything to Merced.
I think the E2k looks interesting to peolple like SGI. CT never posted my translations of the Russian articles, so let me give you a recap. E2K uses a binary recompilator. Something like DEC's FX32!, but with hardware support. Apparently their
recompiler is easily retargetable. They now
can compile x86 and SPARC to E2K and run on the
simulator. They plan to add Merced once the ISA
is fully disclosed. They claim to have booted Windoze on a SPARC and ran Flight simulator, which
supposedly won Babayan a bet with Mr. Ross
(Ross Technologies, creators of HyperSPARC) They can probably add MIPS and any other "legacy" architecture relatively easily...
This is not a limitation of any specific CPU,
but rather of SMP architecture. How many CPU
do you want to hang out on a single chunk of
memory? The ASCI machine with several thousands
MIPS chips was an MPP setup. Sun Enterprise
with 64 CPUs uses a crossbar to memory and god
knows what other tricks that have nothing to do
with CPU itself. You pay dearly for stuff like
this...
well, its not that bad.
build a NOW over Myrinet, just 1.2K per node.
Myricom claims that on PCs, the bottleneck is
PCI DMA performance. Network faster than
internal bus. Not bad. www.myri.com
Of course. Myrinet will set you back $1.2-1.5K
per node in networking gear.
The problem (as far as I understand it) with rc5 on mips isn't that the chips aren't good at it, it's just that too many people have tto much time to optimize Intel rc5 code, and nowhere near as many people are working on the MIPS code.
e;
I saw this a few days ago-pretty cool.
http://micro.magnet.fsu.ed u/creatures/pages/calr12K.html