AMD details warp-pipelining: Triton GEMMs hit 1,437 TFLOP/s on MI355X, matching hand-tuned assembly
xiaosun86 · x · 2026-10-28
AMD's ROCm team published the third post in its GPU engineering deep-dive series on warp-pipelining (block ping-pong): holding two wave groups one stage apart so matrix compute overlaps with data preparation, with barriers maintaining the offset across MI300X, MI350, and MI450.
- On an MI355X FP16 GEMM (M=N=4096, K=8192), an 8-wave warp-pipelined kernel reached 1,437 TFLOP/s versus 1,425 TFLOP/s for a hand-tuned 4-wave kernel — parity, with no register spills in either.
- The technique is exposed through the high-level Gluon block programming model, offering less code, fewer tokens (notably for agents generating kernels), easier verification, and no performance trade-off.
- The post includes stage annotation and barrier examples, shared-memory wait placement guidance, and public kernel examples with commands to reproduce on compute-bound GEMMs.
More from coding & agent
- Ben Horowitz: Databricks' agent system of record lifts accuracy from 50% to 84.5% — irinarish · 2026-10-28
- After auditing his own C code with AI, dev no longer trusts unaudited C/C++ — burny_tech · 2026-10-28
- Snowflake Doubles Down on Agentic RL: Overfit First, Then Scale to 65.7 on SWE-bench — StasBekman · 2026-10-28
- Hook-level agent security wins: Microsoft, Anthropic and OpenAI all shifted from gateways to runtime hooks — jacob_posel · 2026-10-28
- Parsing Claude's Raw SSE Stream into a Forensic Timeline of Agent Tool Calls and Diffs — Additional-Ladder-24 · 2026-10-28
- Supabase launches Compute in private alpha: run AI agents and long-running jobs next to your Postgres data — dshukertjr · 2026-10-28