CUDA Update to FAHCore_22

[foldy]

Will all projects switch to FahCore_22 CUDA on nvidia GPUs or will older projects still run on OpenCL?

[Jonazz]

Everything has switched to CUDA.

[Joe_H]

I don't know if all of the projects using Core_21 have finished. Some were put on hold due to COVID. Core_21 will not be updated to use CUDA.

[foldy]

For performance reason maybe the Core_21 OpenCL will be limited to AMD GPUs? As Core_22 CUDA on nvidia gives speedup.

[psaam0001]

Even if I got a few non-CUDA (OpenMM_21) jobs, I'm ready to run them just the same (on my Windows 10 machine w/the AMD Ryzen 3).

Paul

[Kjetil]

I am not. Core 21 is dead for me, is a old core and is not covid-19.

[Jonazz]

I don't know if all of the projects using Core_21 have finished. Some were put on hold due to COVID. Core_21 will not be updated to use CUDA.

True! All core 22 projects should use CUDA.

[JimboPalmer]

All core 22 projects should use CUDA.

First, AMD and Intel cards do not run CUDA.
Second, Nvidia cards using the Fermi micro-architecture can't use CUDA, so need to be OpenCL
Third, other Nvidia cards should be able to use CUDA, unless there are thermal issues. (over clocking, bad air flow, etc.)

[Jonazz]

All core 22 projects should use CUDA.

First, AMD and Intel cards do not run CUDA.
Second, Nvidia cards using the Fermi micro-architecture can't use CUDA, so need to be OpenCL
Third, other Nvidia cards should be able to use CUDA, unless there are thermal issues. (over clocking, bad air flow, etc.)

Good to know

[toTOW]

Is there a minimum requirement for running Cuda?

Works like a charm on my GTX 1660 Super with Nvidia 450.66; Compute: 7.5; Driver: 11.0
GTX 1660 Super from 730-750k Avg. to 1mil estimated.

On my second Computer with all the old Hardware and Nvidia 390.138; Compute: 3.0; Driver: 9.1
I get the

Code: Select all

Error loading CUDA module: CUDA_ERROR_INVALID_PTX (218)

I don't know if some one already answered this, but the minimum requirements are :
- CUDA 9.2
- Kepler GPU or newer
- no outside toolkit or SDK installed.

[toTOW]

And it is worth mentioning that CUDA is considered as a bonus : projects are still benchmarked using OpenCL to keep things consistent between NV and AMD and to continue using known references without adding one more variable to the process (which is already complicated) ...

[ajm]

And it is worth mentioning that CUDA is considered as a bonus : projects are still benchmarked using OpenCL to keep things consistent between NV and AMD and to continue using known references without adding one more variable to the process (which is already complicated) ...

But, does this "bonus" represent more processing power, more science, or not?
If yes, Cuda should be used and to hell with benchmarking.
If not, there should be no bonus and Cuda should deliver the same PPD as OpenCL.
No?

[NormalDiffusion]

And it is worth mentioning that CUDA is considered as a bonus : projects are still benchmarked using OpenCL to keep things consistent between NV and AMD and to continue using known references without adding one more variable to the process (which is already complicated) ...

But, does this "bonus" represent more processing power, more science, or not?
If yes, Cuda should be used and to hell with benchmarking.
If not, there should be no bonus and Cuda should deliver the same PPD as OpenCL.
No?

With cuda activated you get shorter TPF, meaning you finish your WUs faster. So more science done in the same time...

[foldy]

Is there a minimum requirement for running Cuda?
- no outside toolkit or SDK installed.

That is currently only a workaround until FahCore_22 0.0.14 gets released which ignores other cuda sdk toolkits. Or if you need to have cuda sdk toolkit installed then just remove the global environment variables CUDA_PATH... and only enable them for your other apps which use cuda sdk toolkit

[bruce]

If you're in the business of gathering benchmark data from clients and trying to produce a system that makes optimum assignment decisions, you're stuck with the current fact that Project X or Y can be best assigned to GPU A or B or C based entirely on the measured productivity of that GPU in OpenCL. If GPU B might be able to run CUDA and GPU C cannot, you can't use that information because each project has a single value for productivity. FAH doesn't (presently?) have a way to assign a different set of optimum assignments for devices that can run CUDA than the optimum assignment projects for non-CUDA devices. We don't even have a good way to predict which CUDA-capable-GPUs will end up running OpenCL for a given assignment.

I don't have a good idea how to build a better optimum assignment system than the one that is still being developed -- or whether we really need to anticipate the shortcomings of that new assignment system that's not even ready for a beta test.