Deep systems understanding · Study map 81 of 85
Lab 1 · Explain a single GPU’s limits
Suggested preparation: Capacity and cost: price successful work
Picture it
Imagine setting out to sail across an ocean. Before leaving, you calculate the wind, the currents, and the weight of your ship to predict when you'll arrive. When you inevitably drift off course, comparing your logbook to your initial math is what actually teaches you how to sail. You must explain the gap between theory and reality so clearly that another captain could safely follow your route.
What to understand
- Baseline → memory ledger → roofline estimate → profile
- One-variable sweeps and quality gates
Follow the references
1 · Documentation
vLLM · Serving benchmark reference ↗Use the benchmark interface for reproducible traffic.
2 · Documentation
NVIDIA · Nsight Systems user guide ↗Capture a short trace at the configuration where your estimate and measurement disagree most.
Make it concrete
Deliver a pinned configuration, synthetic workload generator, memory calculation, latency/throughput plots, profiler evidence, and one change that improves SLO-compliant work. If you lack a GPU, finish the memory/roofline worksheet and mark measurements as pending.
You understand it when…
Can you explain the largest gap between predicted and observed latency?
This is a study map. Follow the reference for the full explanation; use the question to test your mental model. Browse the reference shelf.