8-Core Intel Nehalem-EX To Launch This Month

MojoKid writes "What could you do with 8 physical cores of CPU processing power? Intel's upcoming 8-core Nehalem-EX is launching later this month, according to Intel Xeon Platform Director Shannon Poulin. The announcement puts to rest rumors that the 8-core part might be delayed, and makes good on a promise Intel made last year when the chip maker said it would release the chip in the first half of 2010. To quickly recap, Nehalem-EX boasts an extensive feature-set, including up to 8 cores per processor, up to 16 threads per processor with Intel Hyper-threading, scalability up to eight sockets via Intel's serial Quick Path Interconnect and more with third-party node controllers, and 24MB of shared cache."

Comments

March of the penguins

[suso]

Ah! My dream of the day when I can boot up and see penguins taking up the entire screen is almost here.

Re:March of the penguins

[dvlhrns]

AHMEN !!!

Re:March of the penguins

[xmas2003]

If it's penguins you want on your screen:
Gentoo Penguins - King Penguins - Penguin being attacked by a Skua!

Re:March of the penguins

[KillShill]

Penguins and monopolists go together like.... oil and water.

Re:March of the penguins

[MrCrassic]

What are you talking about? I can finally run Vista with that! If I'm lucky, I might even get Aero!

Re:March of the penguins

[Anonymous Coward]

meh... I have an 8-core box now that I've gleefully watched 8 penguins display on while booting linux. Has a pair of quad-core Xeons and 16GB of RAM (I run a lot of VMs in case you're wondering).

Re:March of the penguins

[suso]

Meh? Many of us have already seen 8 and even 16 on a screen, but I'm talking about seeing 32 or 64 on a screen when you have dual or quad chip boards with these 8 core chips.

Probably some people out there have already seen this on non-intel hardware I guess? Like Sparc hardware?

It's obvious

[David Gerard]

Run a REAL operating system, like VISTA!

Re:It's obvious

[natehoy]

Has it been confirmed that you can make a beowulf cluster of these, then? Could we be close to the dream of acceptable Vista performance?

Oh, wait, how many terabytes of RAM can the chipset that runs these handle?

Re:It's obvious

[recoiledsnake]

This processor is meant for servers, because they're Xeon, and with all the Web 2.0 and Cloud computing going on, servers are always hungry for more power.

Re:It's obvious

[ickleberry]

This makes me sad. Web 2-point-Oh is such a waste of a perfectly good 8-core processor.

10 years ago if you had told me about an 8-core processor I would have imagined using it for kick-of-the-ass games, immersive virtual reality, editing 3D video and simulating newer, more deadly designs of chainsaw chain.

But noo, instead they are used to pump out inefficient JavaShit-based versions of the Desktop software we had in '93 with a shiny new rounded corner interface to web browsers around the world. Great.

Re:It's obvious

[StikyPad]

I thought the obvious answer would be octa-porn.

And you're free to interpret that any way you like.

Re:It's obvious

[MrNaz]

Yea, it really bugs me how 95% of a web site's load time and processing load is accounted for by a few pretty features like rounded corners and drop shadows.

How about we put those effects into CSS where they below and not induce massive load by simulating them with 5mb of JavaScript?

Re:It's obvious

[raddan]

Am I the only one here who understands that client-side Javascript has absolutely nothing to do with how many cores your server has?

Web 1.0 can use plenty of cores, too, but generally your Web x.x requirements and your required server core count are orthogonal. Bandwidth and latency requirements for Web 2.0 are a different story, though. Those things tend to scale depending on how shitty your programmers are.

Re:It's obvious

[Sir_Lewk]

Who is talking about servers? I'm thinking about my home machines, you know, where the client-side javascript runs...

Re:It's obvious

[raddan]

http://slashdot.org/comments.pl?sid=1575426&cid=31405814

Re:It's obvious

[Sir_Lewk]

MrNaz was quite clearly talking about clients. If you're new here then you might not have realized it yet, but sometimes discussions actually evolve over the course of several comments.

Re:It's obvious

[Sulphur]

Run a REAL operating system, like VISTA!

Fields of penguins, and Vistas of penguins.

Re:It's obvious

[L4t3r4lu5]

*Hand wave* This is not the CSS you're looking for

It sucks that the implementation is browser-specific, but hey! It's still there.

minimum hardware required....

[rts008]

Now we know what will be needed to run Win 8, I guess.
I better get started on my backyard fusion power plant....;-)

Balance

[Locke2005]

Does it have the memory I/O bandwidth to keep up with the CPUs? When will I be able to actually buy a mother board with 8 of these 8 core CPUs, and what kind of a frame rate would Crysis get on that rig?

Re:Balance

[vivek7006]

Does it have the memory I/O bandwidth to keep up with the CPUs?
Yes
When will I be able to actually buy a mother board with 8 of these 8 core CPUs
When you move out of your parents garage.

Re:Balance

[DarkOx]

These are target it the Virtualization and specialized application space. You are not going to put these in your gaming rig, and your not going to use the +4 core models in your tranditional stand alone application server. You could get much better dollar to performance ration elsewhere if those are your intended applications.

Now slapping two or more of these things on a Linux box with a ton of UMLs running or on VMware ESX, and loading the system up with 128 gigs of ram and a medium business can probalby run their entire datacenter on 2 boxen + an entry level SAN.

Re:Balance

[fuzzyfuzzyfungus]

Given that the Nehalems all have integrated memory controllers, I'd assume that the memory I/O situation wouldn't become substantially worse as you scaled up.

From TFS's mention of "up to 8 CPUs or more with third-party node controllers" I'm(perhaps optimistically) assuming that that means all the RAM in an up to 8 socket system wouldn't be more than one hop away from any core.

They almost certainly didn't go with 24MB of cache because their main memory situation is perfect; but intel's bigger chips are substantially improved from the old "Hey, let's hang a bunch of super expensive Xeons off a dubiously adequate northbridge through a shared front-side bus, let them starve for memory access, and then get curb stomped by cheaper Opterons!" days.

Re:Balance

[Locke2005]

Silly troll! 1) My parents don't currently own a garage. 2) I haven't lived in my parent's basement for over 30 years now. 3) My mother-in-law currently lives in my basement, which might be one of the reasons I can't afford something like this anyway.

Re:Balance

[Anonymous Coward]

Does it have the memory I/O bandwidth to keep up with the CPUs?

Yes. This is what quickpath is all about. It is a genuine, modern, NUMA architecture.

When will I be able to actually buy a mother board with 8 of these 8 core CPUs,

Shortly after launch, if not at launch. Of course, that board will run $2k-$3k, and will only take ECC RAM, but you're not really asking this if that's a problem.

and what kind of a frame rate would Crysis get on that rig?

At lowest settings, I'm guessing about 2FPS. That's because it wouldn't have a video card, because computer cluster machines generally don't.

Re:Balance

[CBRcrash]

I'm thinking computing power for rent (aka the cloud), VDI, cluster data crunching , and any combination of the above

Re:Balance

[afidel]

Basically yes, it actually has more bandwidth for remote NUMA access then current Nehalem-EP systems but fewer memory lanes per core so less bandwidth under high contention or more bandwidth under low contention. IBM has announced the x3690 X5 which has 32 DIMM slots for two EX sockets which will be a killer DB/Virtualization platform if priced competitively.

Re:Balance

[Anonymous Coward]

Does it have the memory I/O bandwidth to keep up with the CPUs? When will I be able to actually buy a mother board with 8 of these 8 core CPUs, and what kind of a frame rate would Crysis get on that rig?

Crysis will get the same framerate as before, because FPS is not due to CPU, its due to GPU.

Go ahead and spend $1000 on a CPU, and i'll spend $1000 on two $500 GPU's with an E8400 or I5. Then we'll see who gets the better framerate

Re:Balance

[Anonymous Coward]

Wait, so you share a room with your mother in law?

Re:Balance

[Anonymous Coward]

Yes... on Friday nights, me, her, and your mom do threesomes.

Re:Balance

[Anonymous Coward]

This is a good point. When do we start crowdsourcing the cloud? Why isn't there a website for this already, where people can buy and sell virtualization services hosted off their local home systems?

Re:Balance

[Locke2005]

Agreed, this is targeted at the DataBase Engine and HPC markets, not at gamers. The frame rate remark was a bad attempt at humor. A much better question would have been, "How long will it take until a massively parallel collection of these gets top ranking on the TOP500 supercomputer list?" I'd give it 2 years.

Re:Balance

[BJ_Covert_Action]

You are not going to put these in your gaming rig

Tell that to the gamers with Epeen insecurities.

Re:Balance

[xenn]

2FPS with NO VIDEO CARD? ...I'm impressed.

Re:Balance

[Hillgiant]

I suppose it is possible to virtualize a CUDA processor in reverse and have, in effect, a low level Nvidia card. Output would be problematic...

Re:Balance

[Anonymous Coward]

a medium business can probalby run their entire datacenter on 2 boxen + an entry level SAN.

I'll give you that you could probably virtualise a hell of a lot of stuff on top of two of these boxes in an ESX cluster (although you'd have no redundancy) but one SAN? You wouldn't have the IOPS.

Sorry, I'm being a pedant...

Re:Balance

[afidel]

Perhaps, it's probably going to be significantly more expensive per raw MIP but because of the higher internode bandwidth it will scale better for many problems so will achieve higher real world throughput. I'd imagine someone will have one placed in the TOP500 by the fall announcement since all the other node interconnect stuff can be pulled from any existing Nehalem-EP based design.

Re:Balance

[Anonymous Coward]

and what kind of a frame rate would Crysis get on that rig?

maybe 20fps?

Re:Balance

[nxtw]

with a ton of UMLs running

UML? Where have you been for the last five years?

KVM or Xen are where it's at on Linux.

Re:Balance

[raddan]

You are not going to put these in your gaming rig

I hear this a lot, but in a modern OS (e.g., one with a good scheduler) and with modern applications (ones that use either threading or cooperating processes), you can easily use a handful of processors, and yes, with normal desktop apps. Google Chrome, for instance, uses the cooperating process model, and for security reasons, I think you're going to start seeing [good] programmers divvy up their applications this way. Not only does it make application security a bit easier (separate address space for each code module), but you get real CPU-level parallelism for free. FreeBSD's new scheduler can even put threads running in the same process on different cores in some cases.

My concern isn't being able to use all those cores-- it's being able to throttle or shut them off when I'm not.

Re:Balance

[Anonymous Coward]

True, though with the lack of mention of SSE, I'm guessing they're not adding any registers. That is definitely a flaw for computational intensity and power conservation, and will hurt them in the competition for HPC clusters.

Re:Balance

[Chris Burke]

From TFS's mention of "up to 8 CPUs or more with third-party node controllers" I'm(perhaps optimistically) assuming that that means all the RAM in an up to 8 socket system wouldn't be more than one hop away from any core.

The block diagram in TFA shows 4 QPI interfaces, so theoretically yes.

In practice almost certainly not, because that theoretical setup has no interfaces left to hook any I/O devices up to it and so is kinda useless. So at least one core has to dedicate at least one link to hook up a PCIe bridge or such. In a large server, you probably don't want all your I/O going through one link (where it then has to fan out from the least-connected CPU to wherever it is needed), so there'll be several CPUs that are not maximally connected.

Even more practically, routing all those links across each other to connect all the CPUs together on the mobo, while also making room for the rather ginormous amount of DRAM channel traces is going to be prohibitive. I'm betting they use no more than 3 QPI links for CPU-CPU interconnect per core, making a 2-hop system (which is still a challenge).

Even MORE practically, the 8-socket market is minuscule, anyone buying into it certainly has a specific application in mind, and you'll get way more bang for your buck making sure your application/OS are NUMA-aware than by trying to minimize hops by doubling the number of layers in your mobo.

That 24 MB of cache is there because it helps even in a 1-socket system, and of their many strengths Intel's greatest is their manufacturing capacity, so it only makes sense to flex that muscle. It's like how they used to get crazy Specfp2000 scores for the Itanium because they could put huge caches on it and that version of specfp should really have been called speccache. Unlike Itanium, though, that giant cache isn't just a layer of polish on a turd.

But in any case, is the Nehalem system architecture "substantially improved" from the P4 shared bus? Oh fuck yes.

Re:Balance

[Anonymous Coward]

Since you're ok with incest, why isn't your wife involved?

Re:Balance

[Locke2005]

I'm sure somebody will be working on one by the fall announcement, but they generally start small and build up over time to the 200,000+ CPUs needed for a high TOP500 ranking. From what people have said on here, this chip does appear designed to do well on a massively parallel Linpack test and many real world MPI applications as well.

Re:Balance

[Locke2005]

Hey! Sex with my mother-in-law is NOT incest! However, sex with the anonymous coward's mom might qualify as bestiality...

Re:Balance

[DigiShaman]

Let's say I invest in solar or wind energy. Which will provide the greatest ROI? Selling the energy directly, or have it run servers to sell off extra CPU cycles?

Re:Balance

[Anonymous Coward]

Hate to tell you this, but Xen development is pretty much over. It's KVM, Virtualbox or VMWare from here on out.

Re:Balance

[daveime]

I was with you right up until "boxen". Then you ceased to exist in my universe.

Re:Balance

[cthulhu11]

I was going to post the very same thought. I'll bet DarkOx is a vi user too ;)

Re:Balance

[joib]

I wouldn't be so sure. Like previous incarnations of the Xeon MP series, this one will be much more expensive per FLOP than the 2 sockets-per-node machines that make up most x86 entries in the top500 list.

Anyway, for these big machines parallel scalability is mostly determined by the internode network, merely stuffing more cores per node does nothing. Or actually, if you don't increase network performance as you make the nodes fatter, parallel scalability will worsen as you have more cores sharing the network link.

Now, one interesting entry that will use these Nehalem-EX chips is the Altix UV by SGI. That will certainly big a very interesting architecture for people looking at big CC-NUMA machines, but as it tops out at 256 sockets in a CC-NUMA configuration it won't get anywhere near the top of the top500 list. Of course, you can cluser together several such machines if your wallet is thick enough, but at that point you lose the global CC-NUMA and a more traditional cluster is more cost effective for MPI jobs.

Re:Balance

[DarthVain]

So what your trying to say is "Never!"

Re:Balance

[initdeep]

it's in the same place that unlimited upload/download 100mbit connections to everyones home are.

Re:Balance

[petermgreen]

Does it have the memory I/O bandwidth to keep up with the CPUs?
IIRC it has quad channel ram so i'd expect the answer to this to be similar to running a quad-core on dual channel ram.

At least according to toms hardware going from dual to triple channel on a quad core doesn't make much difference. so it seems one channel per two cores is about right with current tech.

When will I be able to actually buy a mother board with 8 of these 8 core CPUs
Dunno, I expect you will see the first boards and servers built arround this CPU arround release time (i'm sure the server and motherboard vendors already have engineering samples). I dunno if any of them will be 8-socket though.

and what kind of a frame rate would Crysis get on that rig?
Probablly not that much better than what it gets on current hardware.

Re:Balance

[petermgreen]

I'm(perhaps optimistically) assuming that that means all the RAM in an up to 8 socket system wouldn't be more than one hop away from any core.
These only have 4 QPI links so in a system with more than 4 sockets (remember you have to hook up the IO hubs somewhere) 8-socket system some processors will be two hops away from each other.

there is a diagram of an example 8-core setup at http://www.hardwarecanucks.com/wp-content/uploads/intel_nehalem-ex-8-core.jpg

Re:Balance

[petermgreen]

The block diagram in TFA shows 4 QPI interfaces, so theoretically yes.
Um no, 4 interfaces would only allow five devices to be directly connected.

I'm betting they use no more than 3 QPI links for CPU-CPU interconnect per core, making a 2-hop system (which is still a challenge).
The intel diagram I saw had 3 QPI links on each processor for CPU-CPU (in what was afaict a 2 hop system). Then there were four IO hubs each connected to two processors using the remaining QPI links.

Re:Balance

[petermgreen]

Most server boards i've looked at have onboard video. However this is not shared memory video in the northbridge like you see on desktop boards, it is connected to the southbridge via either 32-bit PCI or PCIe x1 and usually tied in with the remote management stuff. I don't know for sure how well it performs but my suspicion is worse than intel integrated graphics.

Re:Balance

[Chris Burke]

You're right; I meant 2-hop and 3-hop instead of 1 and 2 respectively.

Re:Balance

[petermgreen]

The diagram I saw can be seen at http://techreport.com/r.x/2009_5_26_Intel_unveils_native_eightcore_Nehalem/8p_diagram.jpg

Checking it again it looks like it's a maximum of two QPI hops to get between two processors but a maximum of three QPI hops to get from a processor to an IOH. I agree it looks like a bit of a bitch to route (but then 8 core boards are going to be expensive anyway so maybe high layer counts don't matter so much)

Finally!

[sardaukar_siet]

The end to "can it run Crysis?" jokes!

Re:Finally!

[Anonymous Coward]

Can it run Crysis 2?

http://www.ea.com/games/crysis-2

Re:Finally!

[Cytotoxic]

Even funnier, soon enough you'll be running Crysis on your cell phone (or whatever we call it then). Remember when it was tough to get decent framerate on Doom with high settings? You can run that on a cellphone these days. 15 years from "state of the art" to "runs on my cellphone." Wow. In 15 years you might have a 1TB database running on your personal communicator that fits in your pocket. (in keeping with the "15 years out" prediction theme of the day.

Re:Finally!

[sardaukar_siet]

Dude, do you really think so? I get the whole handheld Crysis thing, but 1TB of data on a personal database? What for? (hey, don't quote me on this, ok?) :D

Re:Finally!

[darthdavid]

3D porn, duh!

Re:Finally!

[sardaukar_siet]

Yeah... or a local mirror of Wikipedia. I get it!

Re:Finally!

[aldld]

But can it run Linux?

Re:Finally!

[kimvette]

Yes, but unfortunately it will only run Crysis on Windows XP. For Vista, you have to wait for the 16-core, I'm afraid. :-/

Re:Finally!

[DudemanX]

By that time I won't need a fake I.D. to rent Ultraporn.

Re:Finally!

[Hurricane78]

It’s a bit bad no say that it always is 15 years. It’s exponential. The goal is growing exponentially. That 15 years it took to raise processing power to where you could run it on a phone, is maybe a couple of weeks nowadays. That gives you a better feeling for it. ^^
Of course the 1TB in 15 years still fits.

Re:Finally!

[DiEx-15]

Yes but... It still won't end the Duke Nukem Forever jokes.

Re:Finally!

[evilWurst]

> In 15 years you might have a 1TB database running on your personal communicator that fits in your pocket. (in keeping with the "15 years out" prediction theme of the day.

Hmm. Applying one of the Moore's Law variants to NAND flash, if storage size for the same price doubles every 18 months, 15 years is 10 generations. 2^10 = 1024. 4-8 GB of flash memory is already relative cheap today, even in the form of a microSD card the size of a fingernail, so I'd be kind of disappointed if we didn't have 1 TB flash drives (or some other tech that eclipses flash) by 15 years from now.

Re:Finally!

[QuantumRiff]

And in 15 years, nuclear fusion will still somehow be 15 years in the future...

programs compatible with 8 cores

[Mantis8]

But how long before game makers and other software companies write code that can take advantage of all those cores? By the time they do, Intel or AMD will have mainstream 32 or 64 core processors on the market.

Re:programs compatible with 8 cores

[Anonymous Coward]

As soon as we start using operating systems that allow the use of all those cores, and as soon as we start using apps where bottlenecks can be multithreaded and spread across many cores.

ie somewhere around 2006 :)

Re:programs compatible with 8 cores

[XanC]

Apache is popular. MySQL is popular. Pretty much any Web or DB server will eat these right up.

Re:programs compatible with 8 cores

[Alastor187]

I am sure there are plenty of applications out there that can take advantage of this new hardware. I run finite element and computational fluid dynamics software at work and both are capable of using the 8 cores in my work PC (dual quad core).

The really sad part though is that for the FEA software I can only use 2 cores because the vendor requires customers to buy a separate HPC license for every processor/core beyond 2.

Re:programs compatible with 8 cores

[hoytak]

Don't know about games, but many types of numerical processing can easily take advantage of this. ATLAS and other high-performance linear algebra libraries already use all available cores (no, IO is often not the biggest bottleneck with these libraries, as they seem to squeeze out all possible advantages from the L1 / L2 caches). In other words, for my scientific computations, I would definitely notice a difference.

Also, OpenMP is becoming easier and easier to use with recent gcc releases, and it only takes a few #pragma statements in some parts of the code to give a huge speedup if you know what you're doing and have appropriate code.

Re:programs compatible with 8 cores

[Wdomburg]

Ever heard of a little thing called visualization? :) Even in the single-host space, there is plenty of software with a high degree of parallelism and horizontal scalability. We have several types of servers that run hundreds (and in some cases) thousands of threads or processes.

Re:programs compatible with 8 cores

[AustinSlacker]

I think you meant virtualization, and yes, that is the target audience for these systems. Been working on them for about a year now and they are quite nice...

Re:programs compatible with 8 cores

[Anonymous Coward]

Comedy gold!

Re:programs compatible with 8 cores

[Klintus Fang]

How long before game makers right code that supports this new chip?
Answer: a very long time. This is a xeon part designed for large database servers. It isn't intended for desktops. Some fool might try to put it into a gaming rig eventually, but that person...really will be worth of the title "fool". That would be like putting an engine designed for a freight train into a ferrari.

When will other software vendors have code that supports this many cores?
Answer: they already do. the companies that write database management software for very large backend database servers already have code that scales to very large core counts. As do many HPC software vendors. That is the intended market segment for this chip and that market segment has lots of software that is ready to burn through all those cores now.

Re:programs compatible with 8 cores

[borgboy]

You're usually better off spending money on spindles and ram and raid controllers way before exotic multicore CPUs, on a database machine.

Re:programs compatible with 8 cores

[XanC]

Not with an X25-E.

IBM Power7 also has 8 cores

[karvind]

If it is matter of core-war, IBM's latest Power7 also has 8 cores. It is actually based on 45nm technology compared to Intel's latest 32nm. What makes Power7 exciting is that it has on-die 32MB L3 cache. They achieved this by introducing eDRAM (embedded DRAM) in the technology. Both Nehalem-EX and Power7 are targeting low-end server market, so it should be interesting battle.

http://arstechnica.com/hardware/news/2009/09/ibms-8-core-power7-twice-the-muscle-half-the-transistors.ars

Re:IBM Power7 also has 8 cores

[blind biker]

Sun's UltraSPARC T1 had 8 cores and a total of 32 concurrent threads, since 4 years ago. Best of all, that CPU is very low-power. Even better, it's completely open-source. You can download everything:
ISA specification
Verilog RTL source code of the design
Verification environment, diagnostics tests and simulation images

Re:IBM Power7 also has 8 cores

[ender-]

I've had 8-core servers for over a year now. Sun T5220s [1 x 8-core x 8-thread] and T5240s [2 x 8-core x 8-thread]. They may not have the raw number crunching ability of the Intel Chips ( and I know nothing about the IBM chips ), but these things can multi-thread like nobody's business! ;) Love seeing the OS report 128 processors!

For the real cpu-hungry stuff - aka. Windows running on ESX :) - we have some 16 core [4 x 4-core] X4450's. I wouldn't mind getting 4 x 8-core Nehalem chips in there.

Re:IBM Power7 also has 8 cores

[maitas]

If EX delivers the expected performance, it will have the same performance per socket than Power7, but half the threads. I prefer the better single thread performance of EX than Power7.

Re:IBM Power7 also has 8 cores

[Hurricane78]

But the specification does not mean anything, since the actual layout on the chip is much harder to get right, than you might imagine.

So since it doesn’t help to get the spec for free: Where can I buy a version that will run my games? (The only resource-intensive processes I run.)

Re:IBM Power7 also has 8 cores

[Anpheus]

Your post is all sorts of confusing.

It is actually based on 45nm technology compared to Intel's latest 32nm.

That makes it sound like more nanometers is better. Not in this case.

Both Nehalem-EX and Power7 are targeting low-end server market, so it should be interesting battle.

You have an extremely interesting definition of low end.

Re:IBM Power7 also has 8 cores

[bertok]

If it is matter of core-war, IBM's latest Power7 also has 8 cores. It is actually based on 45nm technology compared to Intel's latest 32nm. What makes Power7 exciting is that it has on-die 32MB L3 cache. They achieved this by introducing eDRAM (embedded DRAM) in the technology. Both Nehalem-EX and Power7 are targeting low-end server market, so it should be interesting battle.

http://arstechnica.com/hardware/news/2009/09/ibms-8-core-power7-twice-the-muscle-half-the-transistors.ars

Since when is the "low-end server market" made up of 8-core 8-socket machines? Are you from the... future?

2006 called..

[recoiledsnake]

..they want their joke back. Windows 7 runs perfectly fine on 6 year old machines. But MS is known for making shitty OSes with alternate versions so Windows 8 may still suck... though initial impressions are that not much will change from Windows 7.

Re:2006 called..

[toastar]

..they want their joke back. Windows 7 runs perfectly fine on 6 year old machines. But MS is known for making shitty OSes with alternate versions so Windows 8 may still suck... though initial impressions are that not much will change from Windows 7.

every other one sucks? I haven't seen a good windows product at launch since win2k. XP was ok towards the end of it's lifetime. In my mind the jury is still out on 7, It's better then vista, but that doesn't say much. :(

Re:2006 called..

[masshuu]

wait, no, yes, uhh
3.1 = ok(don't have to much experience with it so can't say for sure), 95 = issues, 98 = ok, me fail, xp = ok, vista = Über fail, 7 = ok
so based on that, windows 8 will suck more than Über fail. is that possible? it will be if windows 8 uses a "rent more complicated parts of the OS as you need them" licence, which i have seen people pass the idea around.

Honestly though, unlike vista, windows 7 runs fine, all you need is a couple gigs of ram, which allot motherboards in the past 8 years support. I don't expect you to play crysis on a machine with 1.0 ghz processor, but 7 should run on it.

On a side note, 8 of these in a machine, if i can do quick math in my head, should produce between 2-3 theoretical teraflops of power.

Re:2006 called..

[ooshna]

3.1 was awesome I still remember the days of messing around on it accidentally deleting the file manager and watching my god mothers son spend hours reinstalling first dos then windows with big ass 5 1/2in floppies

Re:2006 called..

[MrNaz]

I use Linux primarily, but I find Win7 to be quite good actually. My experience was kinda like this. So this jury member has decided, and for the apps I use it for (Adobe Premiere/Photoshop) it performs very well, looks pretty and it not too bad from an interface point of view.

Re:2006 called..

[Anonymous Coward]

And this jury member has head from nearly everyone he knows who uses it that its excellent. In one case, someone who failed to get applications from pre-XP working on XP now finds them running in Windows 7 for no reason he can figure out.

Re:2006 called..

[Teancum]

I am glad to see that I'm not a voice alone in the wilderness here on this issue. I also have to agree that the best operating system that Microsoft has ever released was Windows 2000 Professional. It was a significant enough improvement (where it counted.... in the OS core with stability and improved app security and other real OS features, not these crazy and fancy GUI features) over Windows NT 4 and the other operating systems like the MS-DOS based OSs of Win 95, 98, and ME that it really drew a line in the sand in terms of a real substantive operating system to abandon previous offerings to that point in time.

Heck, I'm still using that operating system even now. My only problem is that vendors are no longer supporting this operating system on the philosophy that it is a dead operating system.... and they aren't even trying to support customers even when they have drivers and software explicitly for that OS.

Windows XP seems do be doing OK, but it doesn't really offer much new and in some ways loads up the OS with what I think are useless features I really don't want or need.

On top of everything else, I wish that the developers at Microsoft would also be consistent with their control functions. For users who hare to access the "control panel" to modify settings, it seems like each team of OS developers explicitly decides to completely rewrite the configuration code from scratch and change how to do basic settings.

As a Windows user from Windows 286 and a PC-DOS 1.0 user, I've seen more than my fair share of Microsoft operating systems. I've seen some real duds (MS-DOS 4.0 and Windows 2 are some classic examples) and I've seen some fairly good successes. How Microsoft overcame the duds to make something worthy of attention is to me a miracle in itself.

I'm sort of undecided on the latest offering, but at least willing to entertain the idea. Still, it has much of the baggage from Vista and XP that I love to hate.

Re:2006 called..

[FrankieBaby1986]

Seven is iffy, from a UI point of view. Some of the UI and management changes are nice. Some were just change for change's sake, and quite backward in usability.

Anyone else feel that way?

Re:2006 called..

[billcopc]

Much like game DRM, if Win 8 does that idiotic "rent-a-feature" licensing, someone will come out with full-featured hack before the thing is even released to manufacturing. Illegal ? Sure. Better than the official pay-per-view way ? Definitely. If we are to learn anything from the past, we should know that in most cases the hacked/pirated solution works better than the real thing, because the people designing and building DRM schemes are, by definition, not the sharpest tools in the shed.

Re:2006 called..

[initdeep]

2002 called, it wants it's rant back.

Re:2006 called..

[drsmithy]

Anyone else feel that way?

I don't think any of them were "change for the sake of change", but I do think some of them are definite steps backward for particular types of users (eg: heavy multitaskers like me).

Licensed per Core

[merlinokos]

Software developers are going to have to figure out a new approach to licensing many of their products. VMware, for example, allows you to use a single license for every processor of 6 or fewer cores... how many people are going to pay for another license for the 2 extra cores? I see per core licenses coming in the near future.

Re:Licensed per Core

[Anonymous Coward]

Mainframe software packages already do this. You pay per MIPS.

Re:Licensed per Core

[Anonymous Coward]

I just see retarded, arbitrary licensing restrictions going away.

What if I only want your software to run on 2 of those 8 cores, and I want to use the remaining 4-cores-worth of the license on another box? Oh? You don't allow that? I'll go buy the competition's product.

Or, more likely, you'll just cave in to my demands and sell traditional per-box licenses, regardless of "cores" or "processors" or whatever.

Re:Licensed per Core

[AustinSlacker]

I can certainly see more, not less of this coming soon. With the release of the upcoming 12 AMD processors, having 48 cores on a single system is probably too good to pass up from a licensing standpoint. I am very surprised that no one is licensing by the amount of RAM the app uses. With 1TB of RAM available on a single system, I could see companies licensing that too.
Customer: Your app runs like crap on my server.
Sales Rep: Oh I see you have 256GB of RAM on that system, but in your sales contract, you only purchased the ability to utilize 1GB of RAM. Would you like to upgrade that license? You will definitely see a performance increase!
Customer: WTF??? #$@#@&%
Sales Rep: Cha-Ching!!!

Re:Licensed per Core

[MessyBlob]

Try instead charging for software transactions, which is a useful measure of work done by the software. It's a bit like paying for fuel by the gallon, rather than charging different amounts depending on the size of engine in the car you drive.

Re:Licensed per Core

[nullchar]

That's a cool idea... until the motivation for efficient computing dwindles because the more "user time" the software burns, the more the vendor will charge.

In the case of VMWare, perhaps they could charge you based on the number of startup instances per time period. (But then every time you reboot a VM for maintenance, you get charged, so the incentive for security is diminished. Though one could try ksplice.)

Sun Ultrasparc T2 has 8 cores... and 64 threads

[IYagami]

http://www.sun.com/processors/UltraSPARC-T2/

And the future Ultrasparc T3 will have 16 cores and 8 threads per core for a total of 128 threads per chip

http://arstechnica.com/business/news/2010/02/two-billion-transistor-beasts-power7-and-niagara-3.ars?utm_source=rss&utm_medium=rss&utm_campaign=rss

Re:Sun Ultrasparc T2 has 8 cores... and 64 threads

[Anonymous Coward]

Oh yeah well I'm starting a company up that will have a CPU with 1024 cores. The chip will be more like a chunk and be about the size of laptop.

Re:Sun Ultrasparc T2 has 8 cores... and 64 threads

[Dragoniz3r]

True but they're designed for entirely different workloads. The Niagara series of processors is designed toward large numbers of not-particularly-intensive tasks such as serving web pages and such. Power7 and Nehalem-EX are targeted more toward processing-power-intensive tasks which are still parellizable.

Re:Sun Ultrasparc T2 has 8 cores... and 64 threads

[chris_7d0h]

What the T-series really needs is a boost to per-thread performance since it will otherwise remain a specialty processor only suitable for certain workloads.

The T2 core has more than enough parallelism for most apps out there. What isn't appreciated though is that it pushes the server *implementation details* all the way up to the app-developers, which causes them grief when they need to target different hardware or when they utilize "junior" developers. It also causes a lot longer performance tuning phases than on our previous platforms (SPARC and Intel).

This situation is fine for A-list developers but causes major grief for the multitude of companies whose developers don't have experience in massively parallelized systems. Companies in that situation unfortunately are most out there.

Our local SUN.. I mean Oracle drones always point out that our servers are able to handle so much in parallel, but that means squat if we can't meet our SLAs.

If I have a response time SLA of 1 second, then it does me little good if I can service 10 times the number of requests of a competitor's server if each request takes 3 seconds and the competitor's hardware actually allows me to reach the SLA!

Also, if there is any kind of locking going on, the server will more or less halt and whopty doo there goes the parallelism.

We won't be buying another T2 since event the PHBs can read the productivity charts and risk reports handed to them by external consultants. The cost of performance tuning of apps has climbed a lot and cost us a small fortune and continue to do so. Hardest to cope with in this space are legacy and third party apps where hardly anyone dare update the decayed code (or receive funding to do so) or in the latter case can.

However, we're seriously contemplating buying a bunch of Nehalem-EX servers and would perhaps have bought the Power7 if we were an IBM-shop since both those companies "Get It", get what customers as us need in contrast to Oracle.

Advice to Oracle: Add a bunch of cache and allow for higher clock speed in the T3 to really start competing with Intel and IBM. I don't care if you add a thousand more threads if each thread still incur a latency three to four times longer than your competition.

Finally, yes I realize I come across as a sour grape but the amount of time and cost *wasted* as a result of our PHB buying these servers (based on spec-mark figures) without contemplating the intended workload has really put a dent in my department's work atmosphere.

Re:Sun Ultrasparc T2 has 8 cores... and 64 threads

[Anonymous Coward]

Just out of curiosity, do you know what an Ultrasparc T2 costs? I know POWER6 CPUs cost about $50k for the chip and core activation charges. Compare that to a Xeon that costs about $1k.

Of course the T3's 128 threads aren't so impressive when you consider that it's in-order execution. It has massive memory bandwidth, but it needs it because it has 128 threads! Do you have a workload that has 128 threads? I don't.

dom

can't believe nobody mentioned this by now...

[Anonymous Coward]

"Transcoding pr0n."

When will Moore's Law apply to Cores?

[rberger]

So can we now expect a doubling of cores every 18 months?

Re:When will Moore's Law apply to Cores?

[Colonel Korn]

So can we now expect a doubling of cores every 18 months?

Moore's Law refers to transistor density, right? As long as programming makes the expected shift to massively parallel techniques that would justify a very large number of cores I think the answer to your question is yes.

Re:When will Moore's Law apply to Cores?

[Solapse]

Surely that'll be Core's Law?

Re:When will Moore's Law apply to Cores?

[Anonymous Coward]

As long as programming makes the expected shift to massively parallel techniques

Who is expecting this?

Threading is still a big time sink for programmers, it's easier (cheaper) to throw more hardware at the problem instead of engineering time.
Very few applications even benefit from "massive parallelism" as they're mostly waiting on meatspace or I/O.

Re:When will Moore's Law apply to Cores?

[w0mprat]

Cores themselves had been growing in transistor count even with the increase in cores. So I would say, for now the answer is no. Cores are doing more and more per clock cycle and have more parallelism within the core itself. This is still going on. Overall, specific peformance is still governed by transistor density, Moore's law still holds sway.

Past a point, cores may simpify, I guess thats what we already have in billion transistor GPUs with 1000+ stream processors etc. I would go so far as to say the definition of 'core' is no longer what it once used to be.

Re:When will Moore's Law apply to Cores?

[Klintus Fang]

if you insist on comparing the products that Intel/AMD release into the desktop space with the products that they release into the server space 6-12 months later then yeah. But you are comparing apples and oranges here...

Re:When will Moore's Law apply to Cores?

[rm999]

A core is theoretically a fixed number of transistors, and Moore's law (depending on which version you quote) essentially states that the number of transistors per chip will double every 18 months. Therefore, a corollary to Moore's Law is that the number of cores could double every 18 months. I say could because engineers may decide to put more transistors in each core, which would slow down the core increase.

Also, there is another caveat: many applications will never be able to take advantage of an insane number of cores, so most chips may not follow this route. I actually believe the market for chips will split into two, and that this split has already began. There will be the chips that follow your core-doubling law. These will target gamers (replacing the GPU), scientists, and power-users (photographers, digital artists, etc). The second type of chip will settle on a few cores and will be extremely small/cheap/efficient. These will be the Atom processors of the future, and could provide more than enough power for the typical user.

I believe the cpu demand of most users has mostly plateaued in the past few years (my netbook has the same power as my computer from almost ten years ago), and that is why I wouldn't call chip-doubling a law. The demand for it in the mainstream simply isn't there (at least, yet).

Re:When will Moore's Law apply to Cores?

[hedwards]

Yep, and technically speaking it's only barely related to the size of the chip. It's the number of transistors on a chip. So you can divide those however you like on a single chip and it makes no difference in that respect. For typical home users, I doubt that there's going to be any reason to go with more than a quad core set up any time soon. And possibly ever, the only reason why a typical home user would even want that many is to reduce the possibility of waiting on processor capacity to basically zero.

Home power users on the other hand are probably going to want a combination of a large number of cores and high speed. Expanding need to whatever the high end hardware can do.

Re:When will Moore's Law apply to Cores?

[Targon]

There are some tasks that are very much serial, and these will not benefit from the additional cores, but there are MANY things that can benefit from the additional cores. It is NOT easier or cheaper to throw more hardware at a problem when processors are not getting all that much faster at this point, but are moving to more cores. Really, it comes down to how much understanding of multi-threading the software designers have when they design an application. You will always have some people out there who can't design with running things in parallel, and if they are the ones in charge of the overall program design, then you won't see much of an improvement to adding cores. The thing is, the longer people work on writing multi-threaded applications, the less difficult it will be.

The basic idea here is that if your task CAN be broken up into tasks that can be run in parallel, then adding more cores will improve performance until number of cores = number of threads. This means that quad core, or maybe six cores is the limit for the moment. Now, for games, more cores means that AI can become more complex, or there can be more computer controlled enemies to deal with. In games like The Witcher, a major reason for slowdowns is the number of NPCs that may be wandering around at any given time. If the game were multi-threaded with each NPC getting its own thread, you have a great example where 12 cores would probably be a great improvement. Software rendering, ray tracing and such would also see a benefit to more cores.

It will take some time for old software engines to fall out of use and new one be developed to take their place. Remember, most people DO look for ways to re-use existing code, but a fresh design WILL provide benefits.

Re:When will Moore's Law apply to Cores?

[Kjella]

You just made me realize nobody named Cole (Ashley Cole, Cheryl Cole, Nat King Cole) will ever have a law named after them. Everyone will just snicker and it'll never catch on.

Re:When will Moore's Law apply to Cores?

[Anonymous Coward]

64 cores is more than enough!

Re:When will Moore's Law apply to Cores?

[Bengie]

to go along with what you said, there are more specialized instructions about to come out. One is a new SSE style instruction that is 512bit but can be expanded. They eventually plan on turning this 512bit register into a 1024bit register and will calculated even more parallel data.

There are still more ways to make a single core faster.

Re:When will Moore's Law apply to Cores?

[Anonymous Coward]

Just like a meme.

Hyperthreading

[MobyDisk]

Why are they are still announcing hyperthreading? It was established long-ago that it had no benefit. It's been off on any machines I've ever purchased.

Re:Hyperthreading

[SappoMan]

Because hyperthreading gives you an additional 20/35% of performance in virtualized workloads. Those chips are meant for ESX server farms.

Re:Hyperthreading

[Anonymous Coward]

Depends on what you run on the machine I suppose. I run climateprediction.net on a i7 920 running Linux and get better performance running 8 tasks on 8 cores vs. 4 tasks on 4 cores. Many others on climateprediction.net report the same thing for Linux (don't know about Windows).

Re:Hyperthreading

[Anonymous Coward]

Hyperthreading used to suck due to Intel's initial implementation. The modern versions are decent and should provide a boost. SMT (which Hyperthreading) is in many cases a good/interesting design decision.

http://en.wikipedia.org/wiki/Simultaneous_multithreading

Re:Hyperthreading

[jcupitt65]

Hyperthreading used to suck, but it works pretty well now. In the benchmarks I've done with my code I see about a 60% speedup.

http://www.vips.ecs.soton.ac.uk/index.php?title=Benchmarks#Results_summary

Re:Hyperthreading

[Spatial]

Long ago? CPU architectures aren't static. A Nahalem isn't exactly a Northwood you know.

It can make a big difference to some applications, like 3D renderers. Sometimes it doesn't help, but disabling it without considering the typical load is unwise.

Re:Hyperthreading

[yuhong]

Yep, Intel's optimizations guide has a comparison.

Re:Hyperthreading

[imashination]

Maybe test things rather than just guessing? HT is a ~20% speedup for anyone doing 3D rendering. http://www.cbscores.com/

Re:Hyperthreading

[BostjanSkufca]

According to my server metric graphs the additional threads are only useful for WIO CPU states.

For example, on Intel 4core i7 920 processor, enabling hyperthreading impersonates additional four cores. But CPU utilization reported by metrics software shows that USR and SYS cpu times will only go up to 50% and WIO will add another 12%. This corresponds to having a virtual core used for waiting to IO stuff. Additional 3 virtual cores do not serve anything at all.

Re:Hyperthreading

[Anonymous Coward]

Well written multi-threaded applications see a "hyperthreaded cpu" practically the same as a "non-hyperthreaded cpu". That has been my experience since the first hyperthreading chips first appeared. Poorly written multi-threaded applications suck no matter what kind they run on.

Re:Hyperthreading

[hedwards]

Assuming you haven't disabled it, like you should, for security reasons.

Re:Hyperthreading

[Slashcrap]

Why are they are still announcing hyperthreading? It was established long-ago that it had no benefit. It's been off on any machines I've ever purchased.

Oh God, thank you so much for doing this.

You have given me the confidence to stand up in front of the entire World and scream "I too am a dumb fucking faggot!".

Re:Hyperthreading

[Anonymous Coward]

WoW, never knew this. Looks like I could gain some performance by disabling my hyper-threading!

Re:Hyperthreading

[eabrek]

Why?

Re:Hyperthreading

[Bengie]

HT on the i7. Bit of info.

HT on the p4 sucked because of lack of duplicated units. The P4 had a double pumped integer unit, so HT could dramatically speed up int based calculations, but the FP/SSE was not so if two threads were running at the same time and tried to access the FP, one thread would get the FP and the other would cause a stall. Because of this stall coupled with overhead of switching hardware threads, there was a performance loss.

Jump forward to the i7. Most everything is duplicated. The i7 has a shared SSE, but has separate add/div/mul FP units and duplicated integer units and a few others. This means when HT switches threads, there's a good chance that there will be no stall caused by a contended unit.

A prime example is databases. With the P4, the DB community at large recommended disabling HT. One reason was contended units, but also HT causes the cache to be effectively split, 1/2 for one thread and 1/2 for the other and DBs are cache sensitive. There have been several benchmarks showing the i7 crunching 1.5TB OLAP cubes with results of 200%-400% increase in speed clock-for-clock compared to previous gen Intels(aka core duo line).

Applications that are heavily SSE optimized will more than likely see a slowdown, but this is easily negated by simple optimizations to limit the app from running two threads on the same physical core. In a system/server that has mixed loads, HT could make a dramatic improvement.

I know with my i7, my grid/distributed computing programs claim ~2.7gigaflops per virtual cpu, which means 2flops/cycle per core. That means with HT, I'm getting 1flop per hardware thread. MUCH faster for SSE apps, but only on a per core basis, not per thread. Best bet would be to run 4 threads on SSE and 4 on regular Floats.

Both AMD and Intel can only do 1 DP Float or INT64 per cycle w/o SSE, but Intel with HT can do 2/cycle. So, overly simplified, it's 2xs faster. Most of the time you won't see 2x speed, but you will almost always see a decent amount more than w/o HT. FYI, AMD's future chips coming out next year will have something almost identical to HT.

Crapware

[Archangel Michael]

Now I can run all my crapware, viruses, trojans, malware, and other dubious software bits at FULL SPEED! Yay

I AM....

[Anonymous Coward]

Imagine a Beowulf cluster with these.

AMD's competitor, Socket G32

[yuhong]

In other news, AMD has a blog article on it's soon to be launched competitor to this, Socket G32 8-core/12-core Opterons:
http://blogs.amd.com/work/2010/02/22/magny-cours-is-right-on-schedule-and-shipping-to-customers/

Ditch x86

[MessyBlob]

x86 has been led into a blind alley. Time now for a redesign, to make an instruction set and execution model that doesn't waste 95% of its cycles waiting for memory.

Re:Ditch x86

[fuzzyfuzzyfungus]

I'm pretty sure that x86 doesn't deserve the blame for the fact that on-die cache is astronomically expensive compared to offboard DRAM...

Re:Ditch x86

[Klintus Fang]

People have been arguing as you are that x86's bloated CISC instruction set was inferior to a cleaner RISC architecture for the last 20+ years. Nobody has ever proven that the elegance of the instruction set matters with hard data though.

What evidence we do have goes against that argument.

Apple machines used a cleaner RISC architecture for a while in the desktop space. They never performed any better than equivalent x86 based machines, and in the end Apple abandoned RISC and moved to x86.

Intel came out with a cleaner RISC based instruction set that that the Itanium line uses. If x86 was really as bad as you say, Itanium chips would be running circles around the x86 based server chips provided by both Intel and AMD. That isn't happenning.

Another thing you might not realize: all x86 chips, from both Intel and AMD, once you strip them down to the micro-code level ARE RISC designs under the hood. RISC is the cleaner way to implement the micro code and the underlying execution architecture, but all historical data seems to indicate that the question of whether the instruction set that sits on top of that is RISC or CISC is irrelevant to performance. It is arguably more complicated to design a CISC based chip like x86, but that clearly has not been an obstacle to competing with RISC on the performance end for Intel or AMD engineers.

Re:Ditch x86

[hhw]

People have been arguing as you are that x86's bloated CISC instruction set was inferior to a cleaner RISC architecture for the last 20+ years. Nobody has ever proven that the elegance of the instruction set matters with hard data though.

What evidence we do have goes against that argument.

The only evidence that we have is that the benefits of commoditization and economies of scale often outweigh any architectural advantages. The fact that x86 incorporated many elements of RISC would also demonstrate its value.

Apple machines used a cleaner RISC architecture for a while in the desktop space. They never performed any better than equivalent x86 based machines, and in the end Apple abandoned RISC and moved to x86.

Manufacturing processes simply trumped architectural differences. PowerPC's have never been manufactured on anywhere near the scale of x86.

Intel came out with a cleaner RISC based instruction set that that the Itanium line uses. If x86 was really as bad as you say, Itanium chips would be running circles around the x86 based server chips provided by both Intel and AMD. That isn't happenning.

Itanium is EPIC, not CISC. It is the exact opposite of RISC. It may not be running circles around x86, but that may be due to compilers not yet being advanced enough to take full advantage of the architecture. We may still see this change in the future.

Re:Ditch x86

[Klintus Fang]

You explain away the fact that PowerPC never performed any better than the corresponding x86 counterparts as manufacturing differences yet fail to offer an explanation for why those two are related. Economies of scale mean you can make more chips cheaper. If you think that mass producing something magically makes an inferior product perform better then a product that isn't mass produced please explain how. THere isn't any industry I'm aware of where that happens. You might be trying to argue that PowerPC failed because it didn't have economies of scale and would likely be right. Nonetheless, if RISC was really so much better, it would have been performing better back when Apple was using it and it basically, didn't. No point in arguing about whether Itanium is RISC or not. It has some EPIC features, but it still uses a reduced instruction set architecture at its core. If you chose to not call that RISC, I am curious what you think RISC means.

Re:Ditch x86

[Slashcrap]

Itanium is EPIC, not CISC. It is the exact opposite of RISC. It may not be running circles around x86, but that may be due to compilers not yet being advanced enough to take full advantage of the architecture. We may still see this change in the future.

Have you been in suspended animation for the last 10 years or are you just real fucking dumb? Do you even really know what a fucking compiler is? If you did, you would know that a compiler that advanced will appear at about the same time as better than human AI.

Re:Ditch x86

[MessyBlob]

Although I've been marked as 'troll' (boo!), I'm pleased that some relevant points have been raised in reply. ;o)

I think we're still waiting for a unified breakthrough in core, memory, code, and data design; I can't say what those breakthroughs are, of course.

I'm not complaining about CISC vs RISC, but that our current memory architecture does not serve our instruction sets well, our layered and dynamically abstracted OO code is awkward to implement without queues of indirection, and our compilers are slaves to all the above.

</untroll>

Re:Ditch x86

[Anonymous Coward]

Compare Atom to Cortex.
Using same process, what's the performance, power consumption and die size of each?

Re:Ditch x86

[Klintus Fang]

The problem you are eluding to as I see it is not that the memory architecture does not serve the instruction sets. The problem is that latencies from the core to memory are very long and writing code that is tolerant of those kinds of latencies is either impossible (or impossibly difficult) to do for the majority of typical (desktop in particular) applications. Any code that has a lot of branches (think: any software you write that has a lot of "if" or "switch" statements) invariable ends up reaching the point where the processor cannot perform any more useful work toward running code until it evaluates the condition of the logical expression. And eventually, unless your code fits in the processor's cache, that means stalling and waiting on memory to retrieve some data.

Code needs to be very deeply pipelined to tolerate those kinds of latencies, for some applications that is easy to do, for some it is possible, but very difficult, for many, it simply isn't possible.

Real problem there is that processing speeds in the core have been increasing very fast over the last few decades, and memory technology simply isn't improving at the same rate.

If you look back to the early days (original Pentium was the last time this was true I believe) the speed of the processor and the depth of the processor pipelines actually did match the latency to memory and the cores were slow enough that they were able to absorb a memory latency hit without incurring long multi-cycle stalls. But now the cores are just too fast. Now, if your code needs to wait for a result from memory, its going to wait for a large number of cycles because the cores are so much faster.

No benefit?

[Xocet_00]

This article outlines the various circumstances under which hyperthreading either benefits or impedes performance. While it's true that on average the benefit was zero (meaning about half of what they tested was faster, and about half was slower) there are clearly a lot of applications that see significant performance gains.

It should also be noted that the applications that benefit are ones that would generally be used in Xeon (server and workstation) machines. Further, most of the applications that failed to benefit from hyperthreading are not written to take advantage of many (more than one or two) cores. As applications are updated for "many core" systems, it is likely that the benefit from hyperthreading will become more significant.

In any case, it is far from "established" that hyperthreading has "no benefit."

Re:No benefit?

[MobyDisk]

Okay, perhaps "no benefit" was too strong of a term. There are cases where it helps, and various replies have pointed-out those cases. As a general rule though, if a feature helps 10% of the time, and hurts 10% of the time, I would rather it be off.

Ideally, the feature would be enabled for those apps that do benefit from it. Perhaps the implementation was the problem - the OS general reports twice as many CPUs as it really has, which causes some apps/servers to spawn more threads than they should.

Joy! I hope it comes with ...

[freaker_TuC]

... super cool looking white plastic mold which fits my sochet and cool looking notepad!

Flash

[0123456]

Better than that, with a properly multi-threaded web browser we'll be able to display sixteen animated Flash ads simultaneously with no slowdown!

Well a few reasons

[Sycraft-fu]

First off it went away for a long time. The P4s had hyper threading but the Pentium Ds and Core 2s (duos and quads) didn't. It didn't come back until the i7.

The other reason is that it is useful now. When HT first came out, it was pretty much for desktop chips and we were still very much a single core world. Ok well little was designed to truly take advantage of multiple threads in that environment. People noticed no real speedup. However now not only are things better using multiple cores, but the server market is a target for this as well. On servers, multiple threads per core in hardware work well. You frequently get situations where you have processors that don't need much processor time, but need it often. The context switching can be killer in terms of overhead. More processes on the chip mitigates that can makes more efficient use of the silicon.

Sun is doing this to a much greater degree, in fact. Their new Ultrasparc processors run more than two threads per core. Probably not that useful on a desktop at this point but it can be very useful on a web server.

Hyperthreading is something likely to stick with us at this point. We are moving away from computers that only did one thing at a time, and simply switched back and forth between tasks and towards computers that do a whole lot in parallel.

Somebody's gotta ask...

[unitron]

So, how soon until newegg.com has the fake ones in stock?

Re:Somebody's gotta ask...

[Traze]

Pure awesome!

Re:Somebody's gotta ask...

[weaponsfree]

So, how soon until newegg.com has the fake ones in stock?

For the uninitiated: http://www.overclockers.com/forums/showthread.php?p=6422425

Re:Somebody's gotta ask...

[Dragoniz3r]

Didn't you hear? They're just "demo units"... ;) riiiiiiiight

This again!

[Singularity42]

I get enough hate of x86 in the supermarket tabloids--it's every other story!

statistics

[Anonymous Coward]

Well, we know that's gotta be a lie. He's home on Friday nights. Tuesday night is when he's out bowling.

Cost?

[MSUINTERNS]

I don't see anyone talking about the cost of this, any ideas?

Re:Cost?

[DavidRawling]

100% more than you want to pay and 50% more than you can afford. Look it's for server platforms not a gamer PC. Hyper-V, ESX, Xen, and other types of parallel systems will benefit. SSH won't.

Re:Cost?

[BiggerIsBetter]

I don't see anyone talking about the cost of this, any ideas?

If you have to ask, you can't afford it.

Re:Cost?

[Anonymous Coward]

I have an 8-core, 16GB RAM, w/ 1TB RAID10, Hyper-V server at my house - your point? Some of us have cash to burn. :)

March of the Macs

[fyngyrz]

What could you do with 8 physical cores of CPU processing power?

I already have an 8-core machine, a Macpro, built from two, 4-core MPUs. And I do a lot with it.

Hopefully what this means is Apple will be releasing a 16-core Macpro. Yum! Some saving will be called for, though. [cough]

Re:March of the Macs

[arndawg]

A lot of us have 8 core machines already. But we don't have it on a single cpu socket. No-one cares about your macpro.

Re:March of the Macs

[petermgreen]

I doubt there will be many single socket systems based on this CPU. Remember this isn't on the same socket as any existing CPU and it really isn't designed for such low end systems.

The real news is that this means intel will have a current gen product for machines with more than 8 cores and/or more than 6 ram channels (6 ram channels with 2 sockets per channel* means 48GB with reasonablly priced modules or 96GB with very expensive 8GB modules,).

* in theory you can put three modules on one channel but it limits you to DDR3-800 speeds and none of the boards i've looked at supported it.

Re:March of the Macs

[arndawg]

Did I say a single socket motherboard? no. 8 cores on a single cpu is new (well for x86 at least)...... My workstation has 8 cores, but that's with two xeon 5500s. My point is that his MacPro is nothing spectacular. This CPU enables 16 cores on the same amount of CPUs and is not something his macpro can do.

Fuck it, we're doing 8 cores

[Anonymous Coward]

The onion said it best.

This thing IS a Beowolf cluster

[Joce640k]

all by itself

Looks like its been around for a few months testin

[cloakedpegasus]

http://browse.geekbench.ca/geekbench2/view/198947

Re:Actually...

[vertinox]

It will improve gaming performance if you happened to be running something like Quakes Wars in ray tracing.

Intel put together a demo on a workstation system with two Nehalem quad-core CPUs getting about 15 - 20 fps.

Since ray tracing is embarrassingly parallel, all one needs to do to improve performance is to throw more cores at it.

Keep in mind ray tracing is much more cpu intensive than gpu intensive...

And it's a boutique CPU owned by a software compan

[Anonymous Coward]

Market share for SPARC is roughly on par with Rebbi Moshi's Kosher Pork Rinds at your local deli.

And now SUN has the great good fortune to be owned by Oracle.

We all know what a stellar track record Larry has with hardware right?

What I would do with 8 physical cores, double them

[frambris]

Mmmm.... a bunch of HP DL360s with two of those in each. Yummmmm....

/ Server Nerd

Tell it to MS?

[SmallFurryCreature]

Rounded corners are part of CSS3 and Webkit and Mozilla support it.

Oh and even the largest javascript libraries come in at 100kb, so where do you get 5mb from? Went trolling for it perhaps?

So, next time you complain, check your facts and use a real browser. Not the joke browser that came with your joke OS on your dad's Dell.

I 3 x86

[TheMiddleRoad]

Theoretically... EPIC > RISC > CISC... ...but it doesn't matter much in the real world. The small theoretical speed advantages get lost in the noise of compiler quality, product cycles, bus speed, programmer skill, etc. Underneath, the execution units are similar, and any brilliant breakthrough in EPIC compiler tech will just be copied into the x86 instruction decoder of x86 chips.

So basically, x86 is here to stay for a LOOOOOOOONG time.

Fuck everything, we're doing eight cores

[Anonymous Coward]

Would someone tell me how this happened? We were the fucking vanguard of chips in this country. The Intel Quad Core was the CPU to own. Then the other guy came out with a four core CPU. Were we scared? Hell, no. Because we hit back with a little thing called the Core2 Extreme. That's four cores and a hefty price tag. For the bragging rights. But you know what happened next? Shut up, I'm telling you what happened—the bastards went to four cores. Now we're standing around with our cocks in our hands, selling four cores and an "Extreme" moniker. Extreme or no, suddenly we're the chumps. Well, fuck it. We're going to eight cores.

What's the point?

[Unnamed Chickenheart]

With Supreme Commander 2 being the disappointment it is*, what's the need of 8-core? =-D

* It runs on XBOX. It's dumbed down and it looks and plays like SC&C ( Supreme Command and Conquer). Though the game is not all bad. It seems to be what C&C should have been.

What about Xeon ?

[vikingpower]

One muses whether or not this also is the upcoming end of the Xeon line ?

Re:What about Xeon ?

[petermgreen]

Xeon is brand and nahalem-ex a codename. These chips will be branded as xeon.

The xeon brand has been used for many architectures over the years and I don't see any reason it wouldn't continue to do so just as the names pentium and celeron have been recycled.

What this probably will mean though is that the writing is on the wall for core 2 based xeons now there is a nehalem based xeon that supports more than 2 sockets and more than four cores per socket. I'm sure they will stick arround for a while for legacy support reasons (just as you can still buy core 2 chips for the desktop) but eventually they will be phased out.

Why the same thing we do every night...

[DarthVain]

try and take over the world!