Interesting Stuff on Random Topic

2026.8.29 vLLM-Moet a crazy MoE engine with 2-bit handwritten SASS code (including a reverse-engineer SASS assembler) and SSD MoE offloading. Can serve GLM-5.2 and DSV4-flash on single consumer GPU card (e.g. RTX5090).
2026.8.4 Designing AI Chip Hardware and Software by Bjarke Hammersholt Roune, a detailed writeup from someone who worked on TPU and GPU design.
2025.9.4 A Chinese post from melonedo helpful to understanding swizzling
2025.9.3 Scripts I used to measure GPU L2 bandwidth.
2025.7.3 Cute documentation on layout algebra.