The room smelt of not having been smoked in - Ronald Knox
|
|
|
General purpose GPUs will rule the Earth
GENERAL PURPOSE graphics processing units (GPGPUs) should take the market by storm by the fourth quarter of this year, according to analysts, who reckon this trend will boost AMD at the expense of Nvidia and Intel.
As Moore's Law propels the industry, it is becoming common to look to parallel software tools and applications to tap the potential of the multi-threaded, multi-cored microprocessors in PCs, servers and workstations.
Boffins from market research outfit, GC Research, reckon the age of disruptive parallel computing has arrived and has the potential to change the competitive landscape of the industry in a big way. Not that this should come as a surprise, computer scientists and chip firms have been talking about it for several years.
GC reckons that even AMD's purchase of ATI a few years back goes to prove how vital it was, and still is, for Daamit to push towards a more cooperative form of computing, not only for the benefit of consumer multimedia applications, but also for affordable scientific computing.
With both Win 7 and Apple's Snow Leopard due out on Q409, GC says the industry, and consumer market in particular, can prepare itself for a GPGPU explosion.
GC thinks AMD is best placed to take advantage of this trend, because its engineers have expertise in GPUs and CPUs. Of the three big chip firms, AMD would appear uniquely placed to bridge the world of graphics processors and X86 architecture, say GC's beard-strokers.
Littler chipper, currently somewhat behind rival Intel in X86 performance, is already making strides towards that very goal, with the rollout of its Tigris notebook platform and Fusion marketing campaign. Both efforts put the firm firmly on track to hit its target of having a 32nm-based Fusion platform - integrating the CPU and GPU onto a single piece of silicon - by 2011.
Meanwhile, Nvidia has been somewhat ahead in the GPGPU game, not only offering extra transistors on the GPU for programmability, but also its own Cuda C compiler. The Green Goblin has also said it will support all open APIs, including OpenCL and Microsoft's proprietary Direct Compute X in Windows 7.
But GC reckons that since it's only new apps which can take advantage of Nvidia's current leadership, its GPGPU dominance is actually worth surprisingly little, as the commercial PC space probably won't fully begin to take advantage of the capability for some time.
A transition, says GC, will take years, and this is what will eventually hurt Nvidia, which doesn't have an X86 processor license and thus will never be able to offer an integrated platform. This leaves the firm facing a very distinct possibility of losing entry-level and mainstream PC client platforms to Intel and AMD.
However, GC says that even if Nvidia loses a chunk of the entry-level and mainstream market, its gross margins should improve as its new users will likely be high-end science boffins and professionals.
As for Chipzilla, it too has been chipping away at Larrabee, the high-end discrete graphics engine it hopes will serve as a GPGPU engine. The Intel master plan is to have lower-end desktop and notebook offerings - including a multi-chip packaged CPU on 32nm and a GPU on 45nm - by the first half of 2010 and move the graphics engine to the same process nodes as the CPU by 2011.
This means Intel may well still come out a winner in GPGPUs over the longer term. But GC thinks replacing profitable X86 cycles with cheaper and less lucrative GPGPU cycles may hurt the chip giant's revenue growth a smidgen, especially in the consumer and scientific computing segments.
Meanwhile, Intel continues to add more cores while peddling its multi-threading tech to the masses, just in case developers decide to make apps to suit. µ
Share this:
Share
GPU is blazingly fast only and only for algorithms executing the same
instruction simultaneously upon massive array of numbers - aka SIMD
(Single Instruction Multiple Data). Multi-core Intel/AMD CPU is MIMD
machine (Multiple Instructions Multiple Data) therefore, the
algorithms with inherently diverged code path run more efficiently on
MIMD and SIMD friendly algorithms are better fit for today's GPU. CUDA
interface pretends to be a MIMD but it is VERY missleading, once
threads of warp are out of sync the performance going down to the
speed of single thread - dramatic slowdown. Ray-tracing is one of
examples where MIMD architecture is required to run it efficiently.
The best ray-tracing multi-core CPU implementations outperforms
dramatically the best GPU ray-tracing for big number of triangles and
the bigger number the bigger gap. GPU fans used to compare single
core-2 with GeForce 285 ray tracing and as a rule the CPU code is just
a port of GPU shaders accomplished by authors of GPU ray-tracing
implementations however the optimal for CPU ray-tracing is not optimal
for GPU and vs. verse. Ray Tracing experts knows this GPU/SIMD
limitations too well so NVIDIA worry for its tomorrow 3D positions
where ray-tracing is going to rule. Larabee is going to be kind of
MIMD/SIMD hybrid closer to clean MIMD and G300 is kind of hybrid but
closer to SIMD. Actually even today G200 is not a pure SIMD it has
undependable several SIMD units but still it is really good only for
SIMD friendly algorithms.
Stefan
A GPU and CPU on the same die, making it eventually impossible to get a non-server CPU without GPU. This removes any need for a third party graphics chipset/card, meaning the market for them will die out. Isn't this antitrust stuff?
"disoptimal x86 architecture" ...
well, too many people throw statements in the air like this without any clude of that that are saying ...
Look at the instructions of x86, for example MPSADBW, 8 sum of absolute difference in one instruction in parallel ... or if you look at AVX, it turns that x86 is the most scalable instruction set out there ...
if somebody tell you "disoptimal x86 architecture", he did not do his homework.
Somebody check on Charlie! Jen might have been right and poor 'ol Charlie will be in a hysterical frenzy and might hurt himself!
Yawns... I remember when DEC first intro'ed the Alpha, they provided with VMS, a binary translator. Worked pretty good.
All computer code is a finite set of operations, numeric and logical, hence decomposable and translatable to other formats.
If Intel had had any brains, they would have provided an x86 to Itanium translator.
Oh well, guess we'll be stuck with the crude, disoptimal x86 architecture for some time.
Sylvie,
Many people tried to replace x86 over the history of the PC, and most of them had a compelling story, what they did not have is the compatibility, and the drive to compete as hard as AMD and Intel.
When you look at the strenght of x86, you ll find deep into it the fact that x86 deal with a lot of performance problem automatically. the Out of Order technologies allow "bad code" to perform well, the smart loading units are now dealing with all sort of unaligned load vectors. Assuming that the world will optimize for your hardware is a deadly mistake.
x86 runs all of your legacy code, you can have a 5 years old program, and it will run on your today's computer. This is the main skill of x86 ... and most of the people under estimate this. you can run Vista, or Windows XP ... with outlook 2003 or 2008 ... you can boot MSdos 6.01 ...
x86 is the world of freedom, you are free to mix your software the way you want it.
Legacy is what make x86 unbeatable and required by the consumers, you got to be "compatible".
in comparaison, ARM is the world of Jaillbraking ... because the market is fragmented with little niche markets where the manufacturer decide if you can run Opera or not in your device.
The freedom of x86 world is what make it strong, thinking that a little corner case using the GPU can change this is a very naive thinking, an other marketing hype that will end up forgotten in the long term.
On the top of this, non x86 solution requires drivers to process task ... when you see the list of responsability for blue screen, you know that devices with drivers are not "that trustable" ...
This is my personal opinion, free your mind, run x86!
Running on x86 since 1984!
the hype about market cher in the science sector, this is just flog-blag, nasa effect, check the appleometer. CPUs are cheaper shirley?
Taking their competition with ATI into computeGPUs sacrificed DX10-11 and DDR5 in the now, for GTX300 development later, I cant see that paying off without major programming backup. Doesnt make much sense to me.
Their ARM collaberation should result in multimedia microdevices. A discrete graphics accelerator is not a system on a chip so why are they going that way? Maybe it was a backup plan in case ARM went pear shaped.
Getting between Intel-Larrabee and AMD-ATI is like voyaging between Scylla and and Charybdis. You have to admire their nerve but maybe they prefer a small share of a big pie than a big share of a small one?
I suspect we will also see a comeback of ARM + PowerVR (Imagination Tech) with things like GoogleOS....