DNet Previewing CUDA Client For RC5-72

[ViRGE]

I'm surprised that this hasn't been posted before, but DNet is previewing a build of the DNet client configured to crack RC5-72 on NVIDIA GPUs using CUDA. There are time limited builds over on the pre-release page for Windows, Mac OS X, and Linux.

As for the nuts and the bolts of the matter, it turns out that GPUs are extremely good at RC5-72 (but not OGR) due to the extremely parallel nature of checking keys, even if it's all integer work. The performance difference is simply staggering, my 8800GTX gets over 300MKeys/second, while my QX6850 (3ghz quad-core C2Q) tops out at a bit less than 45MKeys/second (11MKeys/second per thread). Indeed it's so fast that the DNet client has to feed the GPU multiple blocks at once (64 blocks and higher) just to keep the GPU from spending most of its time fetching new blocks.:Q

Anyhow, if you get the chance, DNet is always looking for more beta testers. The stats won't count with the beta client, but this will at least give you the chance to see what your GPU can do once the final client is out.

 

[VirtualLarry]

What happened to RC5-64?

 

[ViRGE]

It was defeated 5 years ago.

 

[VirtualLarry]

Does the DNet CUDA client support multiple GPUs?

 

[ViRGE]

I can't test it, but from what the documentation says, yes.

 

[ViRGE]

I take back what I said about stats; they do count.

 

[Soulkeeper]

this can't be run thru boinc ?

 

[ViRGE]

Nope. DNet uses their own client (which is not necessarily a bad thing, it's very well behaved).

 

[waffleironhead]

Just as an update. Along with the cuda client, there is now a Stream client for all of the ATI owners.

My onboard hd3200 gets ~6,5 million keys/sec
add that to the turion x2 which gets ~7 million keys/sec per core and that is quite the addition to crunching capability.