CODA: Rewriting Transformer Blocks as GEMM-Epilogue Programs | Hacker News

A Hacker News discussion says LLMs can help generate GPU kernel optimizations from ISA docs and other guide-based context.

The thread focuses on rewriting transformer blocks as GEMM-epilogue programs to better fit hardware-specific execution.

Supporters argue frontier models and agents can surface practical low-level ideas for systems programming.

Skeptics warn that subtle implementation mistakes and hidden kernel bugs remain a real risk, especially on new architectures like RDNA 3.5.