The prime-crunching on dedicated hardware FAQ

[jasonp]

Per a post elsewhere, high-end 2009-era GPUs seem to have significant double-precision potential at a cost slightly higher than a bare-bones PC, so at least the cost hurdle is coming down rapidly.

[Robert Holmes]

For the record, the new version of the CUDA toolkit's FFT library now sports a double precision FFT implementation:

http://forums.nvidia.com/index.php?showtopic=102548

Only supported in GT200 cards, i.e. GTX260, GTX280, etc.

[Mini-Geek]

Originally Posted by Robert Holmes

For the record, the new version of the CUDA toolkit's FFT library now sports a double precision FFT implementation:

http://forums.nvidia.com/index.php?showtopic=102548

Only supported in GT200 cards, i.e. GTX260, GTX280, etc.

Might it, then, finally be practical to port Prime95 to CUDA?

[lfm]

Originally Posted by Robert Holmes

For the record, the new version of the CUDA toolkit's FFT library now sports a double precision FFT implementation:

http://forums.nvidia.com/index.php?showtopic=102548

Only supported in GT200 cards, i.e. GTX260, GTX280, etc.

But not ALL GT200 cards. Only the 260 and up. the 250 and 210 don't work. Be very careful.

[hj47]

Originally Posted by Mini-Geek

Might it, then, finally be practical to port Prime95 to CUDA?

This I would LOVE to know.

[Robert Holmes]

It doesn't seem worth the effort --- as said ad nauseam before, double precision on the GT200 sucks: 1 DP unit per SM.

For a top of the line GTX285 card, this gives us (1476*10^6 * 2 * 30) / 10^9 = 88 GFlops, assuming all operations can be turned into MADs. More realistically it will be closer to 40-50, which is what a regular CPU can do. Maybe the GT300 will have proper DP support.

At any rate, a proof of concept wouldn't hurt. I'd be willing to waste some time trying during August, if there was simple enough pseudo-code to begin with.

EDIT:

After a simple experiment, doing a double precision 2^24 Complex to Complex FFT using CUFFT in a GTX260 (theoretical 60 GFlops) takes about 0.0913 seconds, memory transfers excluded. Let's approximate the number of FP operations in the FFT as 5*n log n, i.e. 5 * 2^24 * 24 --- This gives us ((5 * 2^24 * 24) / 0.09) / 10^9 ~ 22 GFlops of FFT. How does this compare to a decent current quadcore?

[Robert Holmes]

Originally Posted by Robert Holmes

It doesn't seem worth the effort --- as said ad nauseam before, double precision on the GT200 sucks: 1 DP unit per SM.

For a top of the line GTX285 card, this gives us (1476*10^6 * 2 * 30) / 10^9 = 88 GFlops, assuming all operations can be turned into MADs. More realistically it will be closer to 40-50, which is what a regular CPU can do. Maybe the GT300 will have proper DP support.

At any rate, a proof of concept wouldn't hurt. I'd be willing to waste some time trying during August, if there was simple enough pseudo-code to begin with.

EDIT:

After a simple experiment, doing a double precision 2^24 Complex to Complex FFT using CUFFT in a GTX260 (theoretical 60 GFlops) takes about 0.0913 seconds, memory transfers excluded. Let's approximate the number of FP operations in the FFT as 5*n log n, i.e. 5 * 2^24 * 24 --- This gives us ((5 * 2^24 * 24) / 0.09) / 10^9 ~ 22 GFlops of FFT. How does this compare to a decent current quadcore?

In comparison, a Core i7 950 running last version of prime95 says:

Timing FFTs using 8 threads on 4 physical CPUs:
Best time for 8192K FFT length: 49.532 ms.

Assuming the linear scaling, current GPUs are no better than CPUs at this, as pretty much everyone expected.

[cheesehead]

Originally Posted by Robert Holmes

Assuming the linear scaling, current GPUs are no better than CPUs at this, as pretty much everyone expected.

But it's not really necessary for GPUs to be better than CPUs in order for a port to be worthwhile, is it?

Even if they're now only in the same range of FFT speed, a port to CUDA could eventually double the potential number of processors GIMPS could use -- assuming one GPU per CPU, and eventually most GPUs are as capable as the now-top-of-the-line models.

[CADavis]

This ^

[lavalamp]

Rather than CUDA, which is for nVidia cards only, might it be worthwhile to write the code in OpenCL or similar?

[nucleon]

Good call not to code for the PS3. Latest PS3 released can't run linux, so by my extrapalation has put up significant barriers to run 3rd party code.

-- Craig