David Kanter (dkanter@realworldtech.com) on 11/9/09 wrote:
---------------------------
>none (none@none.com) on 11/9/09 wrote:
>---------------------------
>>This article:
>>http://www.ddj.com/cpp/207200659
>>
>>has a link to this:
>>http://www.ks.uiuc.edu/Research/vmd/publications/siam2008vmdcuda.pdf
>>
>>At least that is something not coming directly out of nVidia
>>marketing department.
>
>I'm still a little skeptical, since the software stack for the GPU and CPU aren't
>disclosed - although they mention what hardware is used.
>
>Also, UIUC is being funded to do this stuff and is evangelizing CUDA, so it's in
>their best interest to make the GPUs look good.
>
>I'm not saying the numbers are wrong, but I don't know what they represent. You want to know things like:

Well it's easy to quote someone here:

"If you can't speedup those scientific codes by factor 1000, then you really suck as an algorithmic expert"
Dieter Buerssner

Any claim of over factor 10 for any application at this moment is far off reality of course.

In any form of calculation you can do, the GPU's might achieve a factor 5-10 times faster speed. That's not just bandwidth, that's especially because gpu's at this point are simply not faster than that.

It's however obvious that a low level optimized gpgpu program has been programmed by usually a reasonable programmer, whereas the code for x86 usually is total crap.

Most don't even parallellize the software well or not at all in fact. I tend to remember i helped someone out with some quantum mechanics to parallellize it in a tricky manner at 4 cores, moving from 1 core (with tricky i mean that it really would need all 4 cores to be fulltime scheduling the threads to get a good parallel speedup; so if 1 of the cores would be busy with other tasks, it would not scale well).

That also means better cache usage and that you simply suffer less from a bandwidth problem to RAM.

If your chip is good in X and bad in Y, then avoid Y. If chip is good in Y and bad in X then avoid X. That's not what they do with their software however.

>1. What algorithms are used? Do they favor CPU or GPU?

Is it really interesting a claim of factor 100+ is just not realistic for now.

Usually also they compare some oldie P4 cpu with the latest tesla using 4 gpu's and $15k a setup.

The correct compare with that is a 32 core AMD setup with 8 sockets, say 2.2Ghz oldie second hand cheap at ebay available 8354 cpu's and 2 mainboards connected to each other with a lot of RAM.

That's $10k and it eats similar power to a 4 gpu Tesla setup.

>2. What compiler/libraries?
>3. What HW (obviously)?
>4. How well tuned is the app on each stack?
>
>It's really important to know the details when making claims like this. A lot
>of papers that I've seen do a really poor job of this, but they tend to be miles ahead of NV's technical marketing.
>
>In contrast, when I see benchmarks from Intel or AMD on CPUs, they tend to tell
>you how the comparison is done - so you can find out if they are playing games.
>
>David