Building a GPU compiler from scratch: lowering PyTorch to CUDA via six IRs
NoVibeCoding · reddit · 2026-10-13
The author open-sources emmy, a GPU compiler built from scratch, with a GPUMODE lecture and article explaining concepts shared by TVM, Modular, and PyTorch Inductor. Kernel generation is broken into six passes (Torch IR → Tensor IR → Loop IR → Tile IR → Kernel IR → CUDA), each optimized for tasks like fusion or schedule enumeration. Every example is reproducible via a single compile command showing code at any IR level. On benchmarked workloads many kernels outperform Torch Inductor and cuBLAS, using tensor cores and TMA, and the compiler integrates with vLLM for end-to-end serving.
More from Infra
- AMD CEO Lisa Su: agents need different types of compute — firstadopter · 2026-10-13
- AMD's Lisa Su: data center CPU TAM to hit $220B and AI accelerator TAM $1.4T by 2030 — bookwormengr · 2026-10-13
- Dual A100 80GB llama.cpp setup: 50-70 tk/s solo, drops to 10-15 with 2-3 users — vitamins1000 · 2026-10-13
- Telecom stocks plunge in Australia and NZ as SpaceX unveils satellite mobile network plan — Polymarket · 2026-10-13
- A home-server AI that controls smart home, searches web and chats over Matrix — mimi10v3 · 2026-10-13
- Bot traffic dropped after strikes on Yandex data centers — but why? — mobileraj · 2026-10-13