From 3a9c80d33b48f8d1e36fd084ddd8c3f330345372 Mon Sep 17 00:00:00 2001 From: Yunsong Wang Date: Mon, 29 Sep 2025 12:38:12 -0700 Subject: [PATCH] Updates --- docs/best_practices.md | 11 ++--------- 1 file changed, 2 insertions(+), 9 deletions(-) diff --git a/docs/best_practices.md b/docs/best_practices.md index 8cc82a0..f2ae7e8 100644 --- a/docs/best_practices.md +++ b/docs/best_practices.md @@ -127,16 +127,9 @@ Pass: Cold: 0.007819ms GPU, 0.013864ms CPU, 0.50s total GPU, 3.59s total wall, 6 | 63952x | 13.864 us | 432.95% | 7.819 us | 447.95% | ``` -By default, NVBench runs benchmarks on all available GPUs unless specified otherwise. On multi-GPU systems, this can unnecessarily increase runtime and resource usage. To target a specific GPU, saving both time and resources, you can set the `CUDA_VISIBLE_DEVICES` environment variable. In our case, we target the **RTX8000**: - +By default, NVBench executes benchmarks on all available GPUs unless instructed otherwise. On multi-GPU systems, this can lead to longer runtimes and higher resource usage. To focus on a specific GPU and optimize both time and resources, users can use the `-d` CLI option to select the target GPU. In our example, we target the **RTX8000**: ```bash -user@nvbench-test:~/nvbench/build/bin$ export CUDA_VISIBLE_DEVICES=0 -``` - -Now, if we rerun: - -```bash -user@nvbench-test:~/nvbench/build/bin$ ./sequence_bench +user@nvbench-test:~/nvbench/build/bin$ ./sequence_bench -d 0 # Devices ## [0] `Quadro RTX 8000`