Track B — GPU Understanding & ML Performance

Learn to reason about performance: what makes a kernel or a model fast or slow, and how to prove it with measurements. This track turns you from someone who writes kernels into someone who can diagnose and defend performance decisions.

%%{init: {'theme': 'base', 'themeVariables': {'primaryColor': '#F8FAFC', 'primaryBorderColor': '#6366F1', 'lineColor': '#64748B'}}}%%
flowchart TD
  B01["B01 Architecture"] --> B02["B02 Roofline"]
  B02 --> B03["B03 Profiling"]
  B03 --> B04["B04 Numerics"]
  B04 --> B05["B05 Transformers"]
  B05 --> B06["B06 Serving"]
  B01 --> B07["B07 Sharding"]

  classDef trackB fill:#6366F1,stroke:#0F172A,color:#fff
  class B01,B02,B03,B04,B05,B06,B07 trackB

Modules — Track B syllabus (7 modules)

ID	Module	What you build	Status
B01	Accelerator architecture (CDNA vs Hopper)	Runtime device/topology introspection	DONE
B02	Roofline model & arithmetic intensity	An intensity sweep that finds its own ridge point	DONE
B03	Profiling & bottleneck analysis	Four unlabelled "mystery" kernels to diagnose	DONE
B04	Precision & numerics (fp32 → fp8)	Labs that print the hazards; int8 quantization demo	DONE
B05	Transformer performance	Model calculator; per-sub-op layer profiling	DONE
B06	Inference serving optimizations	Batching simulator, KV planner, speculation model	DONE
B07	Distributed training & sharding	Memory planner, collective cost, pipeline bubble	DONE

Suggested order. B01 → B02 → B03 is the diagnostic core: know the hardware, classify the kernel, then confirm with a profiler. B04 covers the numerics you trade for speed. B05 → B06 → B07 apply all of it to transformers, serving, and multi-GPU scale.

Prerequisite: A01 for GPU literacy. B02 is most useful after A02, and B05 builds on A07. See CURRICULUM.md for the full dependency graph.

A note on the tools. Every analytic tool in this track runs on the Python standard library alone — no GPU, no pip install — and labels its output as a model or a simulation rather than a measurement. The lab programs that do need a GPU compute their own numbers; this track publishes no benchmark table it cannot reproduce.