RT opencl on cpu and gpu (vray 2.4.04)

[super gnu]

ive got 2 6 core overclocked machines, one has a gtx 670 in. so seems to make sense for me to use the opencl mixed mode. followed instructions in another thread and installed the amd opencl sdk.. trouble is i cannot see my cpu under the "select opencl device" control. so i uninstalled and downloaded the "intel sdk for opencl apps 2013 r3, cpu runtime" and installed that. now i can see the cpu in the device selection, but on the machine with the gpu i still need to wait several years with my ram pegged at 28gb to get a compile.. so waiting for that now. on the machine with only the cpu, the rt render server recognises the device, but when it starts rendering, prints a load of nonsense on the command line and crashes rt. thats with the intel driver. i then removed it from that machine, and reinstalled the amd driver to see if i missed some setting to install driver.. now the select opencl device control crashes with an error when i try to open it.


great. any suggestions? i suspect remnants from uninstalls maybe piling up.

http://software.intel.com/en-us/vcsource/tools/opencl


http://developer.amd.com/tools-and-s...sdk/downloads/

[vlado]

I guess this is why OpenCL is not too popular :/

I have not tried this with the latest drivers from Intel/AMD, will check them out.

Best regards,
Vlado

[super gnu]

so i left the opencl thingy compiling last night. its done this morning.. however my results are as follows:

2x 6 core machines @4.4 ghz. cpu mode - 1250 kpaths

1x gtx 670 cuda mode - 3890 kpaths

1xgtx 670 opencl mode - 1743 kpaths

as i understood it, cuda and opencl should give similar results?! or is opencl mode not being updated any longer...?

also to be honest im surprised how much faster cuda is than my 2 machines in cpu mode! its not a high end card by modern standards, and its kicking my cpu's ass. is this just why gpu is good, or is there some issue with my cpu result?

[vlado]

How many paths per second do you get with the CPU mode on those two machines?

Best regards,
Vlado

[super gnu]

1250 kpaths

[vlado]

Right sorry, I didn't see that above. OpenCL seems to be really slow on nVidia cards lately; I assume this is because nVidia really wants people to use CUDA...

Depending on the scene though, the CUDA version can be really a lot faster than a CPU render. This is in part because GPUs are made for massively parallel tasks like raytracing, but also because we do a lot of extra stuff on the CPU. The GPU code is a lot simpler and more streamlined.

Best regards,
Vlado