Intel exec says NVIDIA's CUDA will be a "footnote" in history

NVIDIA execs have been talking smack about general-purpose CPUs for a while now, and it looks like Intel's ready to do some talking of its own -- speaking to CustomPC, Intel SVP Pat Gelsinger said that general-purpose GPU computing initiatives like NVIDIA's CUDA would be nothing more than "interesting footnotes in the history of computing annals." According to Gelsinger, the lack of a viable new programming model has held back similarly different architectures like the PS3's Cell because "years later the application programmers have barely been able to comprehend how to write applications for it." That's certainly an interesting point, but we'd say Gelsinger's not really taking stock of the big picture here -- fully utilizing the power of the GPU is the whole point of CUDA, after all, and OS developers like Apple are pushing OpenCL as a way to make GPU acceleration easier to for developers to access. Still, Intel has already said that discrete graphics are on their way out as hybrid tech like Larrabee enters the scene, and Gelsinger basically repeated the party line, saying that and "evolutionary compatible computing model" will be the "right answer long term." Those are some fightin' words -- it's going to be interesting to see how these competing chip strategies play out as other entrants like AMD's Fusion slowly make the scene as well.

[ShadowKain @ Jul 2nd 2008 8:12AM]

Can't Intel and Nvidia just get along? @.@

[initialxy @ Jul 2nd 2008 8:18AM]

it's good to see them fight. competition is good for consumers. especially now AMD is starting to lose the war especially in the CPU market.

[blizz419 @ Jul 2nd 2008 8:36AM]

Dont count AMD out yet, i bet they got somethin up there sleeves that there being super quiet about, i doubt they are gonna just sit there and take a beating without doing something about it, and there really not doing to bad the Phenom black edition is pretty god damn good CPU for its price, its not as good as intels top of the line shit but i dunno about you but i cant afford to spend a grand on a CPU alone, save the loot towards other pricey parts, i mean who's gonna dog a comp with a phenom black that new 128GB OCZ SSD as system drive, some DDR3, ATI's new vid card, i mean that'd be one damn good comp in my eyes yea you can build a better comp but i'd putting those cpu savings into the RAM and SSD is not that bad of an idea

[Richard Lai @ Jul 2nd 2008 1:21PM]

There's also VIA coming along quietly...

[SimbaDogg @ Jul 2nd 2008 4:46PM]

seriously...what the hell happened to the C-E-Oh no you didn't?!?!

[GatgetMan @ Jul 2nd 2008 8:01PM]

Dear Intel:

Don't make me break my footnote off in your ass!

-nVidia

[Dave Chappelle @ Jul 2nd 2008 8:26AM]

To be completely honest i understood very little the first time i read through that article.
Anyways Cuda? i miss the point of this, we already have Intel releasing quad cores with 11ghz overall of power (and only F@H supports 4 cores) then theres the MAC 8 core, so do we need to run our computer of a GTX280 or Quadro card? perhaps i am missing the point of it but it sounds a little secondary to me.

What i want to see is Intel do something like upgrade the gfx cards they release, i am sick of 128mb cards to run 'fifa' or 'madden'

[Kurian @ Jul 2nd 2008 8:39AM]

GPUs can process "certain types" of "small data" calculations hundreds of times faster than any CPU.

You will never see a CPU being replaced by an all in one GPGPU, but you will see GPUs ripping CPUs a new one in certain specialized applications.

[Krush @ Jul 2nd 2008 2:06PM]

@Dave - this is beyond your league man - go play with your cereal box action figures.
"quad cores with 11ghz overall of power"
:D power for a processor comes from how many calculations it can perform per unit time - usually referred to as "flops" (floating point operations per second)

[nikster @ Jul 2nd 2008 6:40PM]

Run Crysis on an 8-core machine without graphics acceleration. Any cheapy graphics card will be better. It's obvious that the computing power in comparatively cheap, massively parallel graphics cards can be used for purposes other than running OpenGL and games.

I think Gelsinger is (A) saying this to protect Intel's interests seeing as it is that Intel has nothing in the GPU market and (B) will end up eating his own words as stuff like OpenCL evolves and provides easy platform-independent access to the GPU power.

I do agree with him that CUDA will be a footnote - an approach like OpenCL - which is not tied to a particular vendor - is a lot more useful.. CUDA was the first step in making programming for GPUs easier, OpenCL is the next one. This isn't going to go away any time soon.

[newcomerX @ Jul 3rd 2008 7:23AM]

Dave, for GPUs it's not just 4 or 8 cores, a little less than a decade ago graphical processors already have that many cores. the number of cores we are talking about here is more than 200. for instance, gtx 280 have 240 cores i think.

if your computing task has a lot can be parallelized, you might achieve more than 10x speedup if not a hundred. you can check out some of their charts, they did achieve 100 time speedup in some of the applications.

But I do agree that the programming interface isn't optimal. it's not so easy to achieve cool speedup if you are not expert.

[Aaron @ Jul 2nd 2008 8:44AM]

When the original Pentium processor came out, Intel was telling the world that ALL discrete components (modems, video cards, sound cards, network cards, etc.) would be eliminated. Why? Because the Pentium processor was powerful enough to take care of all of those functions!

Intel... wake up. There are reasons why houses are not built by one person. Can one person do it? Eventually, yes, but why not ask for assistance in the form of co-processors (e.g., dedicated video chipsets)?

[Thomas Chung @ Jul 2nd 2008 9:22AM]

They already have everything on a chip, it is called System-On-Chip (SOC). In fact, this technology has been around for ages and can be readily found in many embedded device like your super cellphone that can listen to MP3, watch video in HD while getting your GPS information and yet still able to handle all that e-mail and voice calls. Without SOC, you won't have the Razr, or many of the thin multimedia cell phones out there.

With the Pentium, they could have gone down that path. And I'd suspect that SOC for Pentium (with all the subsystems) on the same chip isn't difficult at all because many manufacturers like TI have been doing that for a long time. But integrating all these subsystems requires a lot of budget and in a time, when people are looking for performance, you would want to spend all that budget, resources and transistor count on adding performance to your chip.

[CM @ Jul 2nd 2008 1:26PM]

Owner: Hi, I need a new door for my house. How much is it?
Repairman: That'll be $300,000, we'll replace the house, not just the door.
Owner: WHAT!!!! Why?? I just need the door.
Repairman: You can't. It is one complete package, for your convenience.

This reminded me of a wireless all-in-one printer from HP that I wanted to use to scan a document but I couldn't because it ran out of ink.

All-in-one device (be it pc, printer, phone) is OK as long as it serves its intended purpose right.

[blizz419 @ Jul 2nd 2008 8:46AM]

Macs 8 core is a beast of CPU but its not really a true 8 core CPU i'm pretty sure it's 2 quad core CPU's basically put one chip, when you look at it you can see how it looks like 2 CPU's stuck together side by side

[blizz419 @ Jul 2nd 2008 8:48AM]

that was supposed to be a reply to chappelle didnt seem to work.

[w00t @ Jul 2nd 2008 9:47AM]

I assume you're talking about the 8 core Mac Pro. It uses two standard Intel Xeon Harpertown Quad Core processors on a dual socket mainboard.

It's worth noting all Macs now use off the shelf Intel processors, the same you'd expect to find in equivalent PCs. The only real difference between Macs and PCs nowadays is the software they run and the case they come in.

[blizz419 @ Jul 2nd 2008 9:58AM]

yea just looked back into it and each of those quad cores are really 2 dual cores side by side almost like 2 cpu's glued together, so its really almost like having 4 dual core CPU's not sure if that gives better or worse performance than if it was actually a single 8 core CPU, anybody experts in here that know?

[GatgetMan @ Jul 2nd 2008 8:10PM]

The 8-core Mac Pro is known as such because it has a total of 8-cores. The configuration is a pair of Intel Quad Core Xeon "Harpertown" x5400 series CPUs... The original 8-core Mac Pro housed a pair of quad core "Clovertown" CPUs model x5365....


The current Quad Core Mac Pro has a single x5400 series CPU while the original Mac Pro contained a pair of "Woodcrest" Dual Core Xeon 5100 series CPUs.

...I think I just became a fanboy :-0

[loosely_coupled @ Jul 3rd 2008 9:27PM]

For god sakes, the answer is very simple.....

The Mac Pro is a DUAL PROCESSOR workstation just like many other PC company workstations. They just happen to use TWO quad-core CPUs, given a machine total of 8 cores. But it's disingenuous to call it an "octo-core" as many fanboys like to do, since that would really refer to 8 cores in ONE CPU package.

[Ihar `Philips` Filipau @ Jul 2nd 2008 8:47AM]

Oh, C'mon, Intel.

You did literally the same less than 20 years ago with external FPU.

History's repeating itself.

As it was 15 years ago, people had to maintain two versions of code - one CPU-only and second CPU and FPU - the same happens now: people would (and already do) maintain two (or more) versions of code: one for CPU/plain DirectX, second/etc for nVidia'a CUDA/ATI's Stream/etc.

[Thomas Chung @ Jul 2nd 2008 9:30AM]

One thing we have to consider is programming these devices. While CUDA is Nvidia-specific, as far as I know this is NOT the de-facto API for accessing all GPGPU from all vendors (regardless of whether it is Nvidia or ATI), and picking a choosing one platform over another may become costly in the future in the other one becomes the prevalent one (think Blu-ray vs. HD-DVD). Then the programmer could have kept two versions or three, one for Nvidia, one for ATI and one for the Cell. However, the problem now becomes the maintainability of the code, while some can probably be shared amongst different architecture. I'd suspect that most cannot be kept same across. And we will have to account for the architectural differences across these hardware as well.

This is why I think the x86-larrabee solution is elegant in a way that you are still working with the same instruction set which means that there is very little to change over in terms of porting of code. The other is that now the same code can run on the CPU (with larrabee or not) and it would be transparent to the end-user and even the programmer.

[dagamer34 @ Jul 2nd 2008 10:15AM]

No, it simply means tying yourself to Intel CPUs. The real solution is what Apple is doing, where it has no interest in favoring one GPU manufacturer over another since they must support both.

And this kind of technology really won't take off in PC-land until Microsoft does something about it. No developer really wants to have two sets of codebases to write/debug/maintain. It's a total pain in the ass.

[loosely_coupled @ Jul 3rd 2008 9:33PM]

CUDA or any other vendor-specific tech won't be programmed to by developers, what will happen is a standard "virtual instruction set" abstraction layer
will be created that developers target and which will then converts the instructions to native ATI/Nvidia/Intel's hardware ---- just as was done for graphics programming with OpenGL and DirectX.

Apple's OpenCL system may just become this industry-wide standard...

[tekdroid @ Jul 2nd 2008 10:03AM]

CUDA been great?

[JPitstick @ Jul 2nd 2008 10:29AM]

What happened to CE-Oh no he didn't!?

[Nilay Patel @ Jul 2nd 2008 6:21PM]

Well, dude's not the CEO.

[JPitstick @ Jul 2nd 2008 7:00PM]

@ Nilay- Then come up with something witty for SVPs and other non CEO positions, or just cut the CE-Oh and make it more general like exec_____. I like to know when a fights a brewin.

[Mark Kent @ Jul 2nd 2008 10:35AM]

Shouldn't it be "annals of computing history"?

[wonkydonkydotnet @ Jul 2nd 2008 10:58AM]

Gelsinger says discreet graphics on their way out.
-Obviously the man has never played FPSs.

Besides, if that scientist can throw together 8 PS3s and make a supercomputer, can't somebody get our GPUs ripping?

[Yor1001 @ Jul 2nd 2008 11:10AM]

The big question here, is what performance will larabee be capable of, I mean will it be just another onboard gpu built into the spu or will it be atleast capable to compete with a 8400gt.

[GatgetMan @ Jul 2nd 2008 8:12PM]

Intel's arrogance makes me want to spu...

[thedesolate1 @ Jul 2nd 2008 1:06PM]

This will be great for game consoles to have everything on the GPCPU or what ever you want to call it... Intel is really out of touch with this claim.

[marshall.eisenberg @ Jul 2nd 2008 2:07PM]

Intel has tunnel vision when it comes to the gaming market place - they are a total non player in this segment and have only one motherboard design for this space. In other words, since Intel doesn't participate in the "discrete GPU market space," it can't exist.

They also constantly tout the nonsensical statement of "Intel sells more graphics chips than any other single vendor" in the hopes of convincing people (and themselves) will be content using a crappy generic graphics chipset for FPS games (haha - wotta total joke that is!)

[Lawrence Rodriguez @ Jul 2nd 2008 3:25PM]

Can't we just bring back gladiator fights again and have AMD, Nvidia, and Intel settle it in the Coliseum. AMD gets a short sword and shield, Nvidia gets a triton and a net, and Intel gets nothing since they think they can handle everything with no assistance. Then... release tigers!

[Ethan @ Jul 2nd 2008 4:42PM]

I use integrated graphics, and can confim that they're crap.

[Trent @ Jul 2nd 2008 5:15PM]

GPU functions are going to move onto the die of CPU's. Sony predicted it when they announced Cell and even though I hate sony they were right. AMD has said they will and so has Intel. It's on both their processor maps. Once they get to 8-16 cores per chip both manufacturers are going to use some of those cores for "stream" processors. Essentially devoted GPU like cores that can do massive FPU matrix operations at very high speed. My guess is when 16 cores becomes practical then 8-12 cores will be x86 cores and the rest will be stream processors with hundreds of pipelines per stream processor.

Once AMD and Intel do that Nvidia is out of business because all you need for video can be slipped into the north bridge and would take up very few transistors to provide the frame buffer and signal converters.

[adrian @ Jul 2nd 2008 5:54PM]

Many comments here seem to put down Intel statements, The way I see it, Intel are still running things, Regarding CPU market share.