Why it matters

GPU perf debugging without profiling is guessing. Understanding tools enables real optimization.

Advertisement

The architecture

Nsight Systems: timeline view. CPU threads, GPU streams, kernel launches, memory ops.

Nsight Compute: per-kernel deep dive. Memory, compute, occupancy metrics.

GPU profiling toolsNsight Systemstimeline viewNsight Computekernel deep diveMetricsoccupancy, memory, warpPyTorch profiler + Nsight together give layer-level and kernel-level views
Profiling stack.
Advertisement

How it works end to end

PyTorch profiler: framework-level. Shows layer runtime + memory.

Metrics to check: SM occupancy, memory bandwidth utilization, tensor core usage.

Common issues: uncoalesced memory, low occupancy, tensor core underuse.