Opencl vs cuda

[inoculate86]

If I recall, core 15 was Cuda only. These work units did not require a full core to keep the gpu fed with data. Looking at the CPU utilization of the process was around 0.2% CPU time with some variation but relatively small. Now comparing to the opencl cores, they use a full CPU. Does this mean Cuda has some abstraction that opencl does not have? Specifically, abstract in that allow Cuda apps to move those instructions from the CPU to the gpu?

[bruce]

CUDA is mVidia proprietary so they do their best to sell it over any Open alternative. That includes making improvements to CUDA that never make it into their OpenCL package. That also making the cost of a license which ATI might like to buy prohibitively expensive. ATI has focused on enhancing their OpenCL support. From FAH"s perspective, it's more expensive to have one FAHCore for ATI and another for nVidia, though they did do that for Core_15/Core_16.

The amount of CPU used is governed mostly by the way the drivers are written, whether they're from nV or ATI. Drivers which use 100% of a CPU probably get better frame-rates in games than ones which are more sparing in their use of CPU resources. I don't know if that same distinction also applies to FAH performance or not.

We don't get to vote on the issue, but how much degradation in FAH performance on your GPU would you be willing to accept if SOMETHING could be redesigned to minimize CPU utilization?

[7im]

Like bruce said, I don't think CUDA moved the processing to the GPU, I think OpenCL is doing more processing so the CPU works harder to keep the GPU fed with data. The size of the core_15 work units versus core_21 would seem to support this.

[davidcoton]

@7im
It may not be proven, but is it not likely that the difference in CPU usage is due to different ways of waiting for the GPU to need data transfers? Particular evidence is that, although the nV OpenCL service thread uses 100% of a core, that core can do significant other work without corresponding damage to GPU ppd. It is known that the CPU is only doing useful FAH work when the GPU needs data transfers. Required evidence to support this: Above a certain non-FAHCore CPU load, when the GPU data transfer gets delayed, is the GPU slowed?

[7im]

I agree, the polling method used the driver is also a contributing factor. We've seen similar differences moving through the core progression from _17 thru _21 on NV vs. AMD as well.

[Theodore]

It seems I'm currently folding using CUDA.
At this time, is there a way to disable it, or any updates on how CUDA performs to regular folding?

[Joe_H]

It seems I'm currently folding using CUDA.
At this time, is there a way to disable it, or any updates on how CUDA performs to regular folding?

Currently nobody is folding using CUDA, there is no folding core out that would use it. So you are misinterpreting something that you see in the folding client or otherwise in connection with your system.

[Theodore]

It might be physx ?

[Joe_H]

The current folding cores for GPU folding use OpenCL on both nVidia and AMD based cards. There is a possibility that a future GPU core might use CUDA on nVidia if there was enough of a performance gain to offset the additional development and support cost of maintaining two software codes.

[bruce]

CUDA is proprietary to nVidia so if you have and AMD device, OpenCL is your only option. If you have an nVidia device, you'd like a FAHCore that uses CUDA because performance would increase. As Joe_H suggests, building that second set of FAHCores would be a difficult and expensive task which may of may not be worthwhile to the overall developmental goals of the FAH project.

I'm not sure, but I suspect that running OpenCL on nVidia utilizes important code from both OpenCL which may be passed into CUDA when the functions are the same or similar, but that's all in how nVidia does things inside their driver packages.

Oh, and while you may or may not have drivers for physx on your system, FAH does not use it. It's not worth installing if you have a choice (unless you run something else that does use it).