2026-11-18 · ← News
OpenAI moved AI chip designs from benchmarks into manufactured silicon
OpenAI says its models reduced parts of the Jalapeño chip by 10% to 13%, while AMD adopted changes delivering up to a 14% PPA improvement into a design sent for manufacturing. The important step is that model output survived expert review and reached physical silicon.
The image could not be loaded.
OpenAI has published results from using its models in chip design. For its Jalapeño inference chip, the company says model proposed changes reduced standard cell area by about 13% in the vector unit and 10% in the matrix unit. It says both optimizations were validated during silicon bring-up.
Model proposals made it into chips sent for manufacturing
Work with AMD delivered up to a 14% PPA improvement, combining performance, power and area, over heavily optimized baselines, according to OpenAI. The model rewrote SystemVerilog RTL describing the circuits. AMD included some changes in a design sent for manufacturing and added others to production design files for future versions.
OpenAI also tested five model generations on two open 32-bit RISC-V cores, PicoRV32 and Ibex. At a comparable budget of roughly 112,000 tokens, GPT-6.2 Astra achieved the highest average score on both cores. The full evaluation covered 9,881 scored runs.
Review and verification matter more than the speed of writing RTL
Chip design has a long and expensive feedback loop. A defect that escapes verification may surface only after fabrication. The meaningful measure is therefore the path from a model proposal through simulation and expert review to tape-out, not how many lines of RTL a model can emit.
This also changes the engineer's role. A model can search variants and suggest local optimizations, while people retain the specification, manufacturing constraints and responsibility for correctness. For EDA vendors, the result points toward AI entering the design and verification loop rather than remaining in a chat window.
Company results are not yet an independent benchmark
The production figures come from OpenAI and its partners. The complete methodology, amount of human intervention and failed attempts are not public. PPA gains also depend on the block, libraries, constraints and baseline, so the 14% figure cannot be generalized to arbitrary chips.
The open RISC-V evaluations offer a more comparable measure, but PicoRV32 and Ibex are smaller and easier to inspect than a modern GPU or CPU. What matters is the combination of two forms of evidence: a repeatable evaluation and changes that a partner actually allowed into manufacturing.
Human effort and post-silicon measurements will settle the claim
The next step is to show how much time models save across a complete project, including verification and rework. Reproducible results across more blocks, disclosed success rates and comparisons with the same team working without the model would make the case stronger.
The hardest signal will arrive after fabrication: whether the optimizations repeatedly survive measurement on physical chips without more defects. That will separate a useful assistant for selected blocks from a change in the economics of chip design.
Lilith's verdict
The model is no longer sketching chips on a whiteboard. Its proposal passed engineers and reached the silicon furnace, but the bill for human verification remains behind closed doors.