Inside the MI300X: 23-min visual guide from compute chiplets to a Gluon GEMM
saurabhtwq · x · 2026-11-14
Jino Rohit published Part 2 of "Inside the MI300X," a 23-minute deep dive with extensive visuals and an interactive 3D explorer, building the mental model needed to reason about AMD kernels and read profiling traces.
Key points:
- MI300X stacks 8 compute chiplets (XCDs) over 4 I/O dies; each XCD has 40 physical CUs, 38 active — 8×38 = 304 active CUs
- XCDs hold the CUs, registers, shared memory (LDS) and caches; I/O dies hold the memory controllers managing HBM
- Vocabulary mapping from CUDA: SM→CU, warp→64-thread wavefront, shared memory→LDS
- Ends with a simple GEMM implementation in Gluon
A solid primer for developers writing kernels for AMD GPUs.
More from coding & agent
- Job hunter's trick: fix a real issue in a startup repo, send the PR as evidence — minhnguyen · 2026-11-14
- Screenshot shows multi-agent, multi-user chat with realtime 3D video and long-term memory — nptacek · 2026-11-14
- DeepSeek opens its entire app stack via API: search, voice, OCR, ASR, TTS — churchkey · 2026-11-14
- Garry Tan's 52-task eval: Fable 5.2 matches Opus 5.5 on pass rate but costs 2.80x more — garrytan · 2026-11-14
- YC's Garry Tan open-sources agent harness evals: pass rates tie, costs differ 2.5x — garrytan · 2026-11-14
- T3 Code Now Lets You Fully Customize Your UI In-App; Agent Edits Update on the Fly — yihui_indie · 2026-11-14