If I had a GPU would it help with the NFS project or is this project entirely cpu based? I look at the BOINC stats for this project and the it seems that many people who are the top crunchers for NFS seem to have gpu cards on their computers. Or do they just have many computers that just happen to have GPU's

If GPU computing is no possible what resources would be required to make this possible for the project? Like do you need to hire a programmer or something as I know all things take time and money to do.

Thanks to anyone who can help me with this issue and of course I am new to this computing thing.

NFS@Home is entirely CPU based. Nobody knows how to move the computations that the client performs onto a GPU so that they run faster. This is a research problem, not a programming or money problem.

Its like we're in the stone age when everyone else on the other projects is jumping into light speed...

What other projects do is irrelevant. The client doesn't run on GPUs because nobody knows how to do that.

Take the Sieve of Eratosthenes. The inner loop of this algorithm is very similar to what a GNFS lattice sieve would have to do. An algorithm like this would have a hard time on a GPU because you are performing random updates to a huge array of bytes. GPUs have enormous memory bandwidth but this is a latency problem. Get SOE to run faster on a GPU than a CPU, say 10x faster, and I'll do the same with the client.

So your saying there is hope? That one day if someone can do what you just asked that GPU's could be used or is this just a first step of many to come?

Hey buddy since you know a lot about this sort of stuff what abouthe Sieve of Atkin?

http://en.wikipedia.org/wiki/Sieve_of_Atkin

Also these CUDA guys on this forum talk about what you said too

https://devtalk.nvidia.com/default/topic/415757/sieve-of-erastothenes/

So Im staring to see a little bit of what your saying but I can't pretend for too long so from what it seems we need to be able to implement optimal sieve process on a gpu correct?

Is being ablee to find an optimal sieve to do on CUDA or open CL or any gpu like a holy grail type thing.

Break it down for me if you can there Chief.

Thanks in advance


EDIT: I continued reading and then one forum poster put this thread. You got a point there Jason.





"Memory requirement for bit array is 1250001 bytes.




grid size: 1,1,1

block size: 256,1,1

CPU says there are 664579 primes up to and including 10000000.

GPU says there are 664579 primes up to and including 10000000.

Processing time (GPU): 2201.823975 (ms)

Processing time (CPU): 78.069000 (ms)





Oh no! The CPU is running circles around my GPU. It seems the task is a bit more difficult than anticipated.



My use of global atomics is a real speed killer. But if I don't use atomics, the threads will step on each other's toes.



And I do not yet know how to split up the problem into several thread blocks because I currently communicate any found prime via shared memory, before the "next round" of marking multiples can start.



Well, at least the answer is correct ;) Let's see if my intern can beat me in this little challenge."

SOE is a sieve in one dimension; GNFS requires a rectangular sieve (i.e. two-dimensional).

Honestly, I'm more interested in using GPUs for the postprocessing; this is a problem that using a botnet will not solve.

Thanks for your help. Hopefully one day I'll get a bunch of genius programmers together to gang bang complie a code or do research or some thing to help speed this whole thing it up. I'll call it the Rick James Sieve.

NFS is a helluva drug...

Just a quick note, jasonp is the developer of msieve client used for the post-processing helpers to complete the post-processing stage of the integers sieved here.
So the buddy is a genius.

Yeah but your not full clean unless your zest fully clean.