[H2D/D2H Bandwidth] Bandwidth Degradation Due to Suboptimal One Numa Per Socket

Symptom

Some cards do not meet the requirements in the H2D bandwidth test.

Possible Causes

One Numa Per Socket is not enabled, and four CPUs share one NUMA node. As a result, some CPUs access non-local storage devices, with a slow speed. After One Numa Per Socket is enabled, each CPU accesses one NUMA node.

Solution

Enter the BIOS, choose Memory Config, and set One Numa Per Socket.

For Atlas A2 training/inference products, set it to Enable. For Atlas A3 training/inference products, set it to Disable.