Related Projects

This page compares the whitepaper site and codebase against a few high-signal references. The goal is not to rank them. The goal is to clarify what this repository is trying to teach.

Comparison matrix

Project	What it is strongest at	What this repository borrows	What this repository does differently
CUTLASS	Production-grade GEMM building blocks, template depth, architecture specialization	Respect for explicit tile hierarchy and Tensor Core constraints	Keeps the ladder readable, smaller, and interview-friendly instead of pushing toward industrial abstraction
NVIDIA CUDA Samples	Minimal official examples that clarify APIs and execution concepts	Canonical API usage patterns and baseline correctness expectations	Adds a staged narrative, validation framing, and comparative interpretation
Si Bohm SGEMM worklog	Public optimization diary with strong pedagogical value	The idea that every speedup should be tied to one bottleneck shift	Adds bilingual documentation, explicit validation boundaries, and a more curated research surface
BLIS and CPU GEMM literature	Layered thinking about blocking, packing, and hierarchy	Performance reasoning through memory movement and reuse	Stays GPU-specific, with WMMA guards and CUDA runtime constraints in focus

How to use these references well

1. Use CUTLASS when you want to see where maintainable high-performance CUDA GEMM gets significantly more abstract.
2. Use NVIDIA samples when you want the smallest official demonstration of an API or Tensor Core feature.
3. Use public SGEMM optimization diaries when you want to compare explanatory order and trade-off framing.

What not to confuse

• This repository is not trying to replace cuBLAS or CUTLASS.
• It is not trying to be the smallest CUDA sample.
• It is not a benchmark-only notebook.

Its value is the combination: readable ladder, explicit trust boundary, and a public narrative that can survive technical questioning.