Which forum CUDA programs are DP vs SP?

[VBCurtis]

I am considering the purchase of a secondhand workstation with a Quadro 5000 GPU, which has DP performance = 1/2 SP performance. If I purchase this machine, I surely do not want to waste this GPU running single-precision calculations among my choices at mersenneforum.

Which of our CUDA-enabled programs require DP calculations? Which are SP? I already use msieve poly select, ECM, cudaLLR; I would consider trial-factoring, cudalucas, and perhaps something I'm not thinking of (a sieve for a project I already like, say).

[Mark Rose]

Keep in mind the Quadro 5000 is about as fast as a GTX 960 for LL. It is still reasonably power efficient when doing DP though.

[preda]

TF (mfaktc, mfakto) is SP.
LL (CUDALucas, clLucas) is DP.

LL is also likely limited by memory bandwidth, not only DP performance.

[kriesel]

Originally Posted by preda

TF (mfaktc, mfakto) is SP.
LL (CUDALucas, clLucas) is DP.

LL is also likely limited by memory bandwidth, not only DP performance.

P-1 (CUDAPM1) is also DP. A brief look at the v0.20 source shows lots of variables typed double. Floats are rare by comparison. GPU Memory occupancy I've observed via GPU-Z is P-1>LL>TF. Occupancy limits maximum exponent for a given GPU's memory size, while bandwidth affects execution speed. But it seems reasonable to expect bandwidth and occupancy to be related. Even a 1GB equipped GPU (Quadro2000 for example) can run exponents well above the current Primenet wavefronts.