Can we control CPU core usage in MLPerf on NVIDIA Orin?

Can we control CPU core usage in MLPerf on NVIDIA Orin?

We are running MLPerf on an NVIDIA Orin platform with the following CPU configuration:

• 12 CPU cores
• 1 thread per core
• 3 clusters

Scenario

We tried running MLPerf with different core affinities using taskset. The goal is to limit the benchmark to specific cores and observe the expected core utilization.

Here are our observations:

We expected only the specified cores (e.g., 0-3) to be active during execution, but all cores are being used regardless of the taskset configuration.

Question

Is it possible to control or limit the CPU cores used by MLPerf (e.g., via taskset or other methods) on the NVIDIA Orin platform?

If yes, how can we configure MLPerf to respect the core affinity settings?

Thanks in advance!

Start by reproducing the listed taskset commands on the NVIDIA Orin setup and inspect how MLPerf launches benchmark processes and handles CPU affinity. Determine whether the observed core switching is expected or indicates an affinity issue, then document a verified configuration or limitation with results for the 12-, 4-, and 1-core cases.