Deep systems understanding · Study map 41 of 85
Profiling: find the empty spaces first
Suggested preparation: Rooflines: count work per byte moved
Picture it
Imagine looking down at a complex railway network. A simple 'utilization' metric just tells you the tracks are busy most of the time, but it hides the chaotic reality. A profiling timeline lays out the entire train schedule visually, revealing where trains are stalled at junctions waiting for signals, or where they are speeding freely. By spotting the empty gaps and traffic jams on this timeline, you discover exactly which switch is delaying your network.
What to understand
- Nsight Systems timeline; kernel time versus host time
- DCGM memory/SM activity, power, clocks, throttling, and error counters
Follow the references
1 · Documentation
NVIDIA · Nsight Systems user guide ↗Use CUDA tracing and the timeline views; identify host launches and GPU work.
2 · Documentation
NVIDIA · DCGM feature overview ↗Find profiling metrics and health diagnostics; distinguish activity from useful throughput.
Make it concrete
Capture a short warmed trace. Label the top three time consumers and one idle gap. Propose an intervention tied to that evidence.
You understand it when…
What evidence distinguishes CPU starvation from HBM bandwidth saturation?
This is a study map. Follow the reference for the full explanation; use the question to test your mental model. Browse the reference shelf.