There are 293 ways to make change for a dollar
|
|
|
Intel accidentally praises Nvidia
CHIPMAKER Intel has ended up admitting that Nvidia's cheaper GPU for a supercomputer is at least 2.5 times faster than one of its top of the range chips.
The admission came in a research paper presented at the 37th International Symposium on Computer Architecture.
The big idea of presenting the paper was to knock claims by Nvidia that its GPU was 100 times faster than Intel's x86 processors.
The paper demonstrates that Nvidia's two-year-old GTX 280 GPU is not, in fact, 10 to 100 times faster than the eight month old Intel Core i7 960 CPU at key supercomputing workloads.
It does show that, however it concludes that with the best optimization the GPU is 2.5 to 14 times faster.
While this does pour cold water on Nvidia's claims, it failed to see the real picture of what it was presenting. Namely that a much cheaper GPU chip is faster than Intel's top of the range CPU.
Now we would have thought that anyone trying to build a super computer would welcome a 2.5 to 14 times faster speed. Particularly when it comes at a much cheaper price tag.
So what possessed Intel to prove Nvidia's point?
It seems that Chipzilla was really miffed to see the Green Goblin running around claiming speed boosts of 100 times faster and thought it would take it down a peg or two.
However it totally screwed up and ended up admitting publicly that Nvidia's idea has merit. It would have been better for Intel to walk away and leave the Green Goblin to huff and puff. Instead it hit itself with its own handbag as it tried to lay one onto Nvidia.
We are not the only ones to think that Chipzilla scored an own goal, as you can see here. µ
Share this:
Share
Comparing a CPU to the parallel processing of a GPU isn't really practical in most cases. Many operations benefit from fewer, faster cores and others benefit from a large number of slower 'cores'. There is also the case of the programming language utilised, there are now three different common languages for GPU processing. CUDA (Nvidia only), OpenCL, and Directcompute. Whilst CUDA is more established, its proprietory Nvidia and limits the end user in choice of hardware. Directcompute runs on any graphics chip that supports it but relies on Windows since its related to DirectX. Directcompute is still beneficial though for Windows only based apps and even games. OpenCL is the only true cross platform hardware and software language, and will hopefully be the choice for truly parallel oriented developers. End of this year/early next year will see an evolution of hardware, with the release of Intel's new CPUs (and related sockets), AMD's completely new CPU and sockets, and AMD's (ATI's) new generation GPU. A new system next year, say a new AMD/ATI based system should see a massive improvement over what's currently available and be an enviable platform for parallel processing.
It doesn't matter if the nVidia or ATI GPUs (or GPGPUs) are faster than any x86 processor.
It would cost most companies a fortune to port their software to a new architecture & SMBs/mainstream consumers have no desire to learn and revamp their software.
Ironically, Intel tried this already with Itanium, and when AMD came out with the x86 sledgehammer/Opteron, Itanium didn't make any economical sense and died a rapid death.
If Intel can make the world change their software, what makes nVidia or AMD-ATI think they can???
Now if your dying for a migraine, try compiling your code in a heterogeneous environment (a system with an x86 processor and a GPGPU) to take advantage of both. Good luck with that!
This GPGPU acceleration fad may be fine in an electrical engineering lab workstation... but not much anywhere else.
nVidia doesn't have an x86 license so they have no choice but to tout the performance of their chips because they'll never get any piece of the $trillion dollar WinTel market
If it is "2.5 to 14 times" faster, then that WOULD be somewhere in the "10 to 100 times" faster claim.
Besides, if that's just the 280, the current 480 GPU is surely even faster than that.
Instead of this public nonsense of backwards admitting it charges more for slower chips, Intel should just get some kind of Larabee chip out there.
Maybe they meant 100x the draw from your local power grid...?
Funny, you have total faith in Intel's claim that 100x and greater speed up is not possible on a CUDA platform.
So, if I read your article right, you're happy to accept Intel’s PR, over academic papers with peer review from independent Universities and organisations.
You’re not particularly unbiased yourself are you.
As much as I like the power of the GPU, it is still not the end all solution. This is were Intel fails in omitting pointing that out. For certain applications with few branches it can indeed do a lot of good, but for others it will be completely useless. For instance, nobody is planning on porting the CUDA compiler itself to run on its own target architecture. It wouldn't work very well either.