Re: [AMBER] Multi-GPU Bug in Amber20

From: David A Case <>
Date: Tue, 11 Jan 2022 21:15:54 -0500

On Tue, Jan 11, 2022, James Kress wrote:
>I noticed another oddity in the GPU output section.
>------------------- GPU DEVICE INFO --------------------
>| CUDA Capable Devices Detected: 1
>| CUDA Device ID in use: 0
>| CUDA Device Name: NVIDIA GeForce RTX 3090
>| CUDA Device Global Mem Size: 24268 MB
>| CUDA Device Num Multiprocessors: 82
>| CUDA Device Core Freq: 1.70 GHz
>While trying to benchmark each GPU individually I noticed this apparent
>anomaly in the output.
>I monitored the Amber pmemd.cuda process using nvidia-smi. The GPU 3 was
>the only active GPU. I had set CUDA_VISIBLE_DEVICES=3 and Amber picks that
>up OK. However, the device ID in use by Amber is specified as 0. Shouldn't
>that be 3?

No. The code only sees "visible" GPUS, i.e. those in the CUDA_VISIBLE DEVICES
list. So CUDA Device ID of 0 means it is using the first GPU that is
visible to it. This happens to correspond to device 3 in the list that
nvidia-smi will provide.


AMBER mailing list
Received on Tue Jan 11 2022 - 18:30:03 PST
Custom Search