Amazon's CompoundingKernel: 1000+ agents write GPU kernels from scratch, tops 3 SOL-ExecBench boards
dair_ai · x · 2026-11-18
Amazon researchers present CompoundingKernel, a self-improving agent system that writes GPU kernels purely by search. As of Nov 2026 its kernels rank #1 on the public L1, L2 and Quantization boards of SOL-ExecBench (NVIDIA B200), are 4.40x faster than the reference baseline on 209 public problems, and all 235 problems have site-verified correct kernels.
Key design:
- Agents start from each problem's PyTorch definition only — no reference kernels, no official scores.
- An archive outside the model stores every correct kernel and its lineage; an LLM controller reads it, picks a parent, and gives each parallel child agent one direction. Correctness checks happen outside the model.
- The controller beat a UCB-style tree search on 35 of 38 problems at equal wall time, making 2.4x as many correct attempts; one warm-restart fix raised fleet evaluation rate 3.6x.
Main finding: the rate of reliable measurements — not the number of agents — limits how far such search scales. Over a thousand agents run at once.
More from coding & agent
- GPT-6.2 Astra hands-on: upgrades million-line legacy app, all tests green in 6 hours — thesaraharminta · 2026-11-18
- Dev runs a fleet of coding agents from his iPhone at a conference via FrankenTerm — doodlestein · 2026-11-18
- Runner calling an old copy breaks the fix—absolute launchd script paths solved it — petrusenko_max · 2026-11-18
- Using Opus-4.6 as an LLM frontend to translate other models' output into human-readable form — MilesCranmer · 2026-11-18
- Invisible engineering teams: why DevOps and platform teams get blame but no credit — rseroter · 2026-11-18
- Replit CEO hails Missions for bringing iOS-quality gestures to the web — amasad · 2026-11-18