Inference & Serving · core
GPU Utilization
Compute occupancy, memory bandwidth, kernel launch overhead, tensor parallelism, profiling, and saturation.
Mental model
GPU utilization is not one percentage. Profile kernel occupancy, memory bandwidth, queue gaps, communication, and batch shape to find whether the workload is compute-, memory-, or launch-bound.
How to study GPU Utilization
Begin by restating the mental model in your own words, then connect it to a concrete system you have built or operated. Name the mechanism, the constraint it addresses, and the trade-off it introduces. Use Stanford CS149 Lecture 7 — GPU Architecture and CUDA Programming, Efficiently Scaling Transformer Inference, FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness to check details, but close the source before writing your explanation. Retrieval is the learning step; rereading is only preparation.
Next, compare GPU Utilization with Speculative Decoding, Inference Hardware. Ask what changes in correctness, latency, resource use, operability, and failure recovery. Complete Design exercise: GPU Utilization and preserve the command, input, output, and one failed attempt as evidence. Finish by explaining the idea without jargon to someone who has not studied the track.
Proof of understanding
- Explain the mechanism from first principles and identify the state it reads or changes.
- Give one situation where the concept is the right choice and one where it is not.
- Predict a realistic failure mode before running the drill, then compare the prediction with evidence.
- Connect the result to a roadmap or build artifact instead of treating the concept as isolated trivia.
Learn from primary sources
- Stanford CS149 Lecture 7 — GPU Architecture and CUDA Programming (doc)
- Efficiently Scaling Transformer Inference (doc)
- FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (doc)
- Stanford CS336 — Lecture 5: GPUs (Language Modeling from Scratch, Spring 2025) (video)
- Roofline: An Insightful Visual Performance Model for Multicore Architectures (UCB/EECS-2008-134) (paper)
- NVIDIA Nsight Systems Documentation (doc)
Practice and explain it back
Design exercise: GPU Utilization
Compute occupancy, memory bandwidth, kernel launch overhead, tensor parallelism, profiling, and saturation. Implement designOutline() returning non-empty values for: profile, bottleneckClass, saturationPlan. Each value must name a concrete mechanism or decision.
Expected evidence: A design outline with profile, bottleneckClass, saturationPlan plus an explicit failure mode or trade-off.
Open the interactive drill →Review prompts
- nvidia-smi reports 100% utilization but throughput is poor. Why is that number misleading, and what do you measure instead?
Build evidence
Synthesize: Inference & Serving
Build a production mental model for inference engines, memory, kernels, routing, hardware, and serving economics. Produce one working system, benchmark, or evidence-backed design that integrates the path.
- Implements or precisely models the core mechanisms from all three milestones
- Includes at least one injected failure or adversarial case and demonstrates recovery
- Reports quality, latency, resource, reliability, or usability measurements relevant to the domain
- Ships a concise architecture note explaining decisions, trade-offs, and remaining risks