Vincent Diepeveen (diep@xs4all.nl) on 11/11/09 wrote:
---------------------------
>Potatoswatter (potswa_m@c.com) on 11/11/09 wrote:
>---------------------------
>>Vincent Diepeveen (diep@xs4all.nl) on 11/10/09 wrote:
>>---------------------------
>>>MoTheG (better@not.tell) on 11/9/09 wrote:
>>>---------------------------
>>>>a reader (a@b.c) on 11/9/09 wrote:
>>>>---------------------------
>>>>>stop blabbering.
>>>>>
>>>>>do you have access to GPU nda docs or not?
>>>>
>>>>I do not.
>>>>Does one have to?
>>>
>>>Yes of course,
>>
>>"Of course"? The point of any highly multithreaded machine is to shift the focus
>>from instruction latency to memory latency. NV gives a rough idea of the operations
>>and their latencies and a better idea of the memory architecture.
>
>Let's face it potatoswatter, the only successful software on GPU's right now has
>been programmed by NCSA type guys, be it the chinese organisation equivalent or
>the US version of it. Only those guys have lots of information that's not on the
>net, or have the tools and toys, and have them long before they ever got released.
>
>That's the reality.
>
>All other hard working amateurs at home, yes even total top coders, they all didn't
>manage to outperform quadcore hardware with ANY gpu; of course provided that their x86/x64 software was coded real well.

On what problem?

I generally agree with you. GPU hype is out of hand and SIMD+multicore is way underrated as a buzzword and an optimization platform.

But saying "0.2 IPC" without context is just as meaningless as their "200x speedup". On the face of it, NV doesn't claim (at least I'm not aware of a claim) their product to be suitable for problems where the number of threads is at all limited. IPC is a useless measure for an SMT multiprocessor designed to saturate an uncached memory channel. It's not a secret that full-width multiplication is slow, 0.25 IPC, on NV80. It's assumed that any developer accepts that and is prepared to fight it with ever-more threads.

As for the lack of successful CUDA programs, the learning curve to get them to work seems beyond all but the most tortured grad students. Then, the driver is too unstable and unfriendly for home use. Finally, serious development doesn't want to be tied to NV.

But on the other hand, the members of http://gpugrid.net would probably tell me to eat *my* words.