GPU Usage

[Brain]

I'd like to find out how many people out there do GPU work, especially for PrimeNet. As far as I know, there are mfaktc, CUDALucas, llrCUDA and genefer (all CUDA) and early trials in OpenCL. PrimeNet accepts the first two. So here's the poll:

[Brain]

80% CUDALucas, 20% mfaktc. The latter mainly for score hunting.

[vsuite]

Initially only mfaktc
Now CudaLucas mainly (since they take so long to finish) and short bursts of mfaktc (sometimes running same time as CudaLucas, sharing the GPU).

[Christenson]

GEForce 210 (cheap, not bought for GIMPS) on a dual-core Pentium 4, doing only mfaktc, doing 53M exponents to depth 72. Runs about 5M factors per second, haven't found out how to make it coexist well with normal console use (writing programs, editing, browsing, running autocad viewer) on that machine, so mfaktc doesn't run if I'm doing work on it. I report results once or twice a week to GIMPS.

Bunch of other non-GPU machines ATM, 2 at home including 6-core Phenom II, 4 at UVA of various vintages plugging away at ECM, P-1, LL-D, and LL. One more, about 5 years old at home, need to get xubuntu running on it, have been too distracted to get done. All report to GIMPS, though NSF@home might start getting some cycles as they factor such numbers as M1061.

TF no longer makes sense except on a GPU.

[xilman]

My gpu-usage falls neatly into two categories: polynomial searching for GNFS in my leisure activities and CUDA software development in professional life.


Paul

[ET_]

Mfaktc, CUDALucas and ppsieve

[Honza]

Now only TPSieve; some CUDALLR and GeneferCUDA earlier.

[diep]

Originally Posted by Brain

I'd like to find out how many people out there do GPU work, especially for PrimeNet. As far as I know, there are mfaktc, CUDALucas, llrCUDA and genefer (all CUDA) and early trials in OpenCL. PrimeNet accepts the first two. So here's the poll:

Brain this is impossible to figure out statistics on gpu computing in a commercial manner; those who have software well to work at gpu's, they really don't want to make any noise and profit from it bigtime. Furthermore AMD/ATI and Nvidia really obfuscate what happens there as well, let me politely call it incompetence driven by greed.

If they know anyhow what happens over there; if a company orders a couple of thousands of gpu's then obviously they're not going to spam entire planet for that. Relatively seen price is so low. What is it, if you buy a 1000 at a time, you're looking at under half a million dollar; compare a single 8 socket box Xeon already is a similar price when fully configured.

The only 'victims' of this is the open source community so to speak, at which also many public scientists depend; they have no clue what already happened there past decades in gpu computing.

So the real successtories you will never hear about, as that has direct consequences for the persons involved. GPU programming in itself is not new. Already somewhere start 90s there were GPU type arrays you could buy for not too much, which for specific type of work total outgunned any supercomputer.

The arrival of bigger caches lately and the huge number of cores and huge power budgets now have made it attractive to do gpuwork at home for the hobbyists.

Please note more and more applications slowly start to use it as well, yet because of the obscurity and big greed in which things have been going, it really doesn't get used as much as i thought it would happen.

It's the big commercial companies with unlimited need for computational power that already use them for years. There is very few jobs though where they openly write that they need a GPU expert to program their stuff.

Vincent

[TheJudger]

mfaktc (not really a surprise, is it?) Most of the time I'm running -pre releases (last stage of my test procedure).

Oliver

[Christenson]

@diep:
I don't think Brain has any issues with the limitations of his poll, which go well beyond what you identified. Think "informal poll" and "rough guesstimate"...not "precise mathematics"