Advanced

Profiling with Nsight

Nsight Systems shows where an application spends time; Nsight Compute investigates individual kernels. Use the timeline to choose what deserves detailed profiling.

This concept is part of Premium, which includes all concepts, course lessons and answers.

Get Premium · ₹2,000 / $256 months of access · one payment · no auto-renew

RELATED CONCEPTS

Memory Coalescing Occupancy and Register Pressure Kernel Fusion Roofline Model torch.compile and CUDA Graphs Stragglers and Hangs

LESSONS THAT TEACH THIS

Kernels and the MachineWrite a reduction, then explain why the naive one is slow18 min→Kernels and the MachineFusion removes traffic that tiling cannot19 min→Kernels and the MachineThe roofline turns 'it is slow' into a number and a next step23 min→

PRACTICE THIS IN REAL QUESTIONS

GPU & Accelerator ArchitectureAn Nsight summary reads DRAM 90%, SM 30%, occupancy 45%, L2 hit rate 20%. What is the bottleneck?→CUDA, Triton & Kernel EngineeringA training step runs at 20 percent model FLOPs utilization. Profile it and find where the missing time goes.→CUDA, Triton & Kernel EngineeringHere is an Nsight Compute report for a slow kernel. Read it, name the bottleneck, and tell me what you would change.→LLM Inference & Servingnvidia-smi shows 30 percent utilization on our serving fleet. Is that a problem, and what would you look at instead?→CUDA, Triton & Kernel EngineeringExplain FlashAttention. Why is it called IO-aware, and what does it actually save?→CUDA, Triton & Kernel EngineeringExplain occupancy and register pressure: launch bounds, spills, the calculator, and why 50% occupancy can beat 100%.→