CUDA apps very slow on FX 570?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi!

I wanted to try some CUDA enabled applications and discovered that there must be something wrong with the current driver (or my Quadro FX 570 is damn slow…).

My configuration:

Vista Business SP1, 32-bit

3GB RAM

Intel Core2Duo Quad 2.4GHz

Quadro FX 570 with 256MB RAM on PCI Express x16

Quadro driver 182.08

To start testing CUDA I’ve tried to run the bandwidthTest from the CUDA-SDK. The output is:

Host to Device Bandwidth for Pageable memory

.

Transfer Size (Bytes)   Bandwidth(MB/s)

 33554432			   1523.5

Quick Mode

Device to Host Bandwidth for Pageable memory

.

Transfer Size (Bytes)   Bandwidth(MB/s)

 33554432			   1317.3

Quick Mode

Device to Device Bandwidth

.

Transfer Size (Bytes)   Bandwidth(MB/s)

 33554432			   5923.8

Don’t know if those values are withing range but as a second test, I tried the BarsWF MD5 bruteforcer (0.8 CUDA) and the results are quite disappointing: My CPU computes about 150 million hashes per second while the GPU only reaches 25 million hashes. I expected the GPU to be several times faster than the CPU!

Other CUDA apps I’ve tried also perform unexpectedly slow (ElcomSoft’s Password Recovery Agent for example)…

Am I doing something wrong here?

2

The Quadro FX 570 does not appear to be a very fast GPU. It has only 16 stream processors (not sure what the shader clock rate is) and the device-to-device memory bandwidth, as you found, is very low.

For comparison, the GTX 280 has 240 stream processors clocked at 1.3 GHz, and has a device-to-device bandwidth of at least 114 GB/sec.

3

Thanks for making this clear. So I can give up fixing possible driver problems, can I?