OpenAI's Jalapeño chip: AI helped design it in 9 months, then tuned kernels to 88.94% of peak
创业邦 · wechat · 2026-09-18
OpenAI has disclosed first results for its in-house inference chip Jalapeño, which went from initial design to fabrication in just 9 months with an internal team averaging under 100 people, while Broadcom handled much of the physical design and manufacturing work.
AI played a central role in development. The team used Google's XLS toolchain to translate chip design into software-like problems AI models handle well, running a feedback loop of engineer-set constraints, model-generated code, and tool verification to speed up speed/power/area tradeoffs. After first silicon arrived in May, internal models optimized programs running on the chip: an attention kernel improved from 0.31% of theoretical peak to 88.94% in about 40 hours. Using AI coding tools, the team also brought three open-weight models not in the original production plan to high performance within two months.
The deeper shift: chips are now being designed for AI programmers too—keeping architecture clear and behavior predictable so AI can continuously write kernels and tune performance, cutting the long-term software maintenance cost of custom silicon. Deployment is planned to begin by year-end.
Related event: OpenAI's first in-house AI chip Jalapeño taped out in 9 months(5 posts)→
More from coding & agent
- Claude Rebuilds the Iconic Earthrise in 3D, Down to the Second — Simple-Diver-2192 · 2026-09-23
- Hexis Shares How They Manage Agent Skills Permissions on Top of Git — juanviera23 · 2026-09-23
- Avelune: A 3D Multiplayer World Simulated at 20Hz on Vercel with an AI NPC — evilrabbit_ · 2026-09-23
- LLVM-Lens: Open-Source Tool Visualizes the Full LLVM Compilation Pipeline — sh4dy_0011 · 2026-09-23
- No single model to rule them all: developer juggles Opus, GPT-6 Astra and Chinese models as agents — iannuttall · 2026-09-23
- How to run Qwen3.8-27B with 160k context on a 16GB AMD card: full config — According_Study_162 · 2026-09-23