Eighteen sessions with concrete outputs

Plan for 45–90 minutes per session, three sessions per week. If a prerequisite is new, repeat a session; the exit task matters more than the calendar. Keep your answers and measurements in a copy of JOURNAL.md.

Week/session	Work	Command or file	Produce before advancing
1/1	Read orientation and draw CPU→GPU flow	Lesson 0; python -m learning.check --list	Definitions of host, device, tensor, launch
1/2	Trace a partial block, then implement offsets	python -m learning.check offsets	Passing checks and N=17/BLOCK=8 diagram
1/3	Map slices and transposes to storage	python -m learning.check strides	Six physical indices for the exit question
2/1	Build the sum tree	python -m learning.check reduction	All levels for width 5 and an explanation of padding
2/2	Work softmax with three numbers	python -m learning.explain softmax	Hand-calculated max, exponentials, denominator
2/3	Implement and debug stable softmax	python -m learning.check softmax	Passing extreme-logit and shift-invariance checks
3/1	Work residual RMSNorm by hand	python -m learning.check rmsnorm	Per-row/per-column/per-element labels
3/2	Implement rectangular matrix multiplication	python -m learning.check matmul	Correct 2x3 @ 3x2 product, no aliased output rows
3/3	Trace K tiles and output edge masks	python -m learning.explain matmul	Diagram for M=5, N=7, K=5 and toy tiles
4/1	Check GPU environment; complete Triton add	python -m learning.check triton_add	Passing tails in FP32 and FP16
4/2	Complete Triton row sum	python -m learning.check triton_sum	Passing strided-row case and dtype explanation
4/3	Complete Triton softmax	python -m learning.check triton_softmax	Passing full-row output and stability cases
5/1	Calculate timing units and Amdahl's law	python -m learning.check measurement	Correct bandwidth and overall-speedup example
5/2	Predict and benchmark softmax warp counts	kernel-bench --op softmax --shape 1024 1024 --dtype float32	Prediction, JSON path, observed medians
5/3	Compare graph replay and ordinary dispatch	Same command with --timing events and a different output file	Explanation of the changed measurement boundary
6/1	Compare eager and compiled RMSNorm	kernel-bench --op rmsnorm --compile	One carefully scoped result sentence
6/2	Read CUDA softmax and inspect GEMM PTX	cuda/kernels.cu; python scripts/inspect_kernel.py --op matmul	CUDA cooperation diagram and one identified PTX instruction
6/3	Reimplement one kernel from a blank file and explain it	Journal	Ten-minute walkthrough and one honest unresolved question

Notes on the order:

• Lesson 7 comes in week 5, not between lessons 6 and 8. Its CPU exercise is paired with the GPU benchmarks that use it (sessions 5/2 and 5/3), so you calculate speedups and GB/s right before measuring your own.
• Session 4/1 starts with the one-time GPU setup (on Windows, .\scripts\setup_windows.ps1). If installation takes the whole session, that is normal. Finish triton_add in the next session and let the rest of the schedule slide by one.
• Weeks 1–3 need no GPU. The optional "GPU step" commands in lessons 5–7 belong to weeks 5–6.

What each part of the schedule demonstrates

Question	Session evidence
Can you reason about memory correctly?	Offset ownership, tail masks, row gaps, strides
Can you derive an operation?	Stable softmax and RMSNorm handwritten calculations
Can you write and debug a kernel?	Your three edited Triton exercises and named failing/passing checks
Can you explain hardware tradeoffs?	CUDA reduction cooperation, GEMM input reuse, measured warp-count experiment
Can you make a credible performance claim?	Same-shape/dtype baselines, graph versus dispatch, saved raw results

The beginner path prepares foundations; compiler implementation and distributed collectives remain additional topics.