- Documentation
Volcengine Kubernetes Engine
Volcengine Kubernetes Engine
FAQs
Heterogeneous computing FAQs
Why does the process SM utilization not match the computing power proportion configured for the mGPU?
Heterogeneous computing FAQs
Why does the process SM utilization not match the computing power proportion configured for the mGPU?
Why does the process SM utilization not match the computing power proportion configured for the mGPU?
Issue description
The process Stream Multi-processor (SM) utilization that is viewed through the NVIDIA Management Library (NVML) API or nvidia-smi pmon does not match the computing power proportion configured for the mGPU.
Solution
The SM utilization collected by official NVIDIA tools indicates the SM utilization within a fixed cycle. The computing power proportion configured for the mGPU indicates the proportion of computing power resources used by user applications throughout their lifecycle. These two metrics are measured on different time dimensions and cannot be directly compared.
To evaluate the computing power isolation capability of mGPU, you need to measure it by using performance metrics, such as QPS, of the user application.
Last updated: 2026.07.07 19:32:58