Phase 3 · The Metal

Matmul Mastery

The rite of passage: chase cuBLAS.

0/3 · 430 XP

You’re previewing the curriculum. Sign in (top right) to track progress, take knowledge checks, and unlock the verifiers.

Tasks

Work through Simon Boehm's CUDA matmul worklog
Naive → coalesced → tiled → vectorized → warp-tiled, with measured speedups. The de facto kernel-interview prep (its author got hired to Anthropic's perf team).
read+100 XPresource ↗
Pass the grader: tiled matmul at ≥40% of cuBLAS
Shared-memory tiling in CUDA C++ or Triton. Graded on correctness plus measured TFLOPS as a fraction of cuBLAS on 4096³ — GPU-agnostic by construction.
kernel+250 XPauto-verified
PMPP patterns: reduction, scan, histogram/atomics
Chapters 9–11 (4th ed. numbering) + GPU MODE lecture 9. Reductions underlie softmax and every norm.
read+80 XP