Hiding/Disabling one GPU in a multiGPU system

[SandboChang]

Background: ubuntu 18.04, ROCm 2.1
GPUs: Vega FE2, Radeon VII1

Would there be a way to hide one GPU in the system?
I am trying to run some tests with the GPUs, but many of the benchmark scripts do not allow selecting a particular GPU device. Thus if I can hide the GPU from the software it will be very helpful.

With Nvidia's card, it seems like this can be done by "CUDA_VISIBLE_DEVICES" to hide a GPU from a test. Would there be a similar feature in ROCm?

[jlgreathouse]

We make this avaialble as a language-level feature rather than a ROCm-wide feature. So depending on the programming language your benchmark uses, you will need to set a different environment variable:

• OpenCL: GPU_DEVICE_ORDINAL is a comma-separated list of the devices you would like to be visible to your OpenCL-using application.
• HIP: HIP_VISIBLE_DEVICES is a comma-separated list of devices you would like to be visible to your HIP-using application.
• HCC: HCC_DEFAULT_GPU can be used to automatically change which GPU HCC will use for GPU offload.

[SandboChang]

We make this avaialble as a language-level feature rather than a ROCm-wide feature. So depending on the programming language your benchmark uses, you will need to set a different environment variable:

• OpenCL: GPU_DEVICE_ORDINAL is a comma-separated list of the devices you would like to be visible to your OpenCL-using application.
• HIP: HIP_VISIBLE_DEVICES is a comma-separated list of devices you would like to be visible to your HIP-using application.
• HCC: HCC_DEFAULT_GPU can be used to automatically change which GPU HCC will use for GPU offload.

Thanks a lot for your detailed reply. I should have been more specific, I am trying to run some benchmarks which were provided here:
https://github.com/tensorflow/benchmarks/tree/master/scripts/tf_cnn_benchmarks

So it is using Tensorflow with ROCm. In this case, I am not sure if it falls into the category of OpenCL/HIP/HCC. Appreciated if you can let me know.

[jlgreathouse]

AMD's TensorFlow implementation uses the HIP language runtime, so you should probably use HIP_VISIBLE_DEVICES.

[SandboChang]

Thanks, by adding this line HIP_VISIBLE_DEVICES=2 in front of python3 ....benchmark...., I was able to run it on the desired GPU.