Gemini 4 Neon autonomously writes GPU kernels, 5.9x faster prefill on AMD hardware
sebkrier · x · 2026-11-22
An AMD engineer gave Gemini 4 Neon Thinking Machines' Inkling 2 Mini GGUF and let it handle the entire optimization on a Strix Halo (8060S) GPU: writing kernels, profiling, finding bottlenecks, and benchmarking autonomously. It worked 4.3 hours with 490 tool calls, beating llama.cpp on ROCm:
- Decode: 44.8 tok/s vs 23.8 (1.88x)
- Prefill: 1,465 tok/s vs 233 (5.9x)
- 4-stream aggregate: 100.8 tok/s vs 54.1 (1.86x)
Generated kernels reached 80% of the GPU's theoretical memory bandwidth. A model autonomously writing and optimizing low-level GPU code to this level is a striking demo of agent capability in systems engineering.
More from coding & agent
- Solo consultant's 11-tool AI stack: Claude Code builds, Codex reviews, bots run marketing — jacob_posel · 2026-11-22
- Turn your AI edits into a reusable writing style guide with one prompt — hwchase17 · 2026-11-22
- Neon agent turns photos into Blender 3D models via colmap in one hour — eigenhector · 2026-11-22
- Letter of Agency: Signed, Expiring Credentials to Stop Agents Getting Blocked — moss_in_it · 2026-11-22
- Claude reports reaching "a genuine pause point" after ultra-long session with codex as task master — craigbalding · 2026-11-22
- On-Chain Audit of 8,000 x402 Endpoints: 89% Valid, 542 Token-Domain Mismatches — JonahDiedrick · 2026-11-22