How to diasble Tensorcore FP32 in A100 when running HPL included in NGC?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello.

I’m now running HPL in NGC HPL-benchmark image (21.4-hpl) on DGX-Station A100.
The initial results of HPL seemed to use the tensorcores in A100 for HPL application.
It’s performance was over Rpeak of FP64, that means that the A100 utilized the tensorcores while calcuating HPL (It’s my guess).

Here is my question.

How to disable the Tensorcore FP32s in A100 for running HPL in NGC HPL-benchmark image (21.4-hpl) on DGX-Station A100?
Is there any external control (eg. envronmental variables) for stopping the tensorcore FP32?

I want to get the results by using FP64 only on DGX-Station A100.

Best regards.

2

There isn’t any way to do that. And the calculation mode is not FP32. A100 Tensorcore have a FP64 path.