Nsight Compute
A GPU kernel-level performance analysis tool provided by NVIDIA. It measures a single kernel's throughput, occupancy, DRAM read/write bytes, and memory throughput as a percentage of peak, helping determine whether a kernel is compute-bound or bandwidth-bound.
Timeline (last 90 days)
Under some system configurations (especially with Secure Boot enabled or when running in virtualized/container environments), profilers such as Nsight Compute may be unable to access GPU hardware performance counters, causing profiling to be limited or to fail, and the error messages are often not intuitive enough.
Hardware performance counters are dedicated registers inside the GPU used to tally metrics such as SM utilization, memory bandwidth, and instruction issue counts, and Nsight Compute reads these counters through driver interfaces (such as CUPTI).
Nsight Python lacks the interactive source-level correlation view that Nsight Compute provides.
Nsight Python is better suited to being embedded in existing test or tuning scripts for quick screening and regression comparison, and after anomalous kernels are found, graphical tools can be used to dig deeper.
Using Nsight Systems in conjunction with Nsight Compute can pinpoint whether a performance bottleneck is memory-bound or compute-bound, enabling the selection of an optimization direction.
The number of GPU hardware performance counters is limited, and once dcgm-exporter occupies them, tools such as Nsight Systems and Nsight Compute cannot simultaneously use the same set of counters.
All Facts (6)
Under some system configurations (especially with Secure Boot enabled or when running in virtualized/container environments), profilers such as Nsight Compute may be unable to access GPU hardware performance counters, causing profiling to be limited or to fail, and the error messages are often not intuitive enough.
50%UnverifiedHardware performance counters are dedicated registers inside the GPU used to tally metrics such as SM utilization, memory bandwidth, and instruction issue counts, and Nsight Compute reads these counters through driver interfaces (such as CUPTI).
50%UnverifiedNsight Python lacks the interactive source-level correlation view that Nsight Compute provides.
50%UnverifiedNsight Python is better suited to being embedded in existing test or tuning scripts for quick screening and regression comparison, and after anomalous kernels are found, graphical tools can be used to dig deeper.
50%UnverifiedUsing Nsight Systems in conjunction with Nsight Compute can pinpoint whether a performance bottleneck is memory-bound or compute-bound, enabling the selection of an optimization direction.
50%UnverifiedThe number of GPU hardware performance counters is limited, and once dcgm-exporter occupies them, tools such as Nsight Systems and Nsight Compute cannot simultaneously use the same set of counters.
50%