← The feed Many Minded · The brief
KNOWLEDGE · forward · impact 2/5 · 2026-11-18 · OpenAI

OpenAI says its models are now trimming real silicon, including AMD's

OpenAI reports model-written chip changes that shrank area in its Jalapeño inference chip and improved AMD designs by up to 14%.

OpenAI published research on 18 November 2026 describing how it uses its models to improve chip design. It began using GPT-4o in early 2025 to help design Jalapeño, its first custom inference chip. The team estimates model-proposed changes cut standard-cell area by about 13% in the vector unit and about 10% in the matrix unit. One multiplier redesign used roughly 15% less cell area and dropped latency from two clock cycles to one. These optimizations were silicon proven during bring-up.

With AMD, OpenAI reports up to 14% improvement in power, performance and area over heavily optimized baselines on AMD's flagship GPU IP, by rewriting SystemVerilog RTL. OpenAI calls this about a year of progress in under 5 months. AMD put some changes into a recent design sent for manufacturing and added others to production files for future versions. AMD also verified single-digit PPA gains across key blocks of its upcoming flagship CPU. Synopsys tools checked that rewrites preserved behavior and measured the trade-offs.

OpenAI also benchmarked models on two open-source RISC-V cores, PicoRV32 and Ibex, across 9,881 scored runs. GPT-6.2 Astra scored highest on optimization, and in the from-specification task it reached about 0.7 on PicoRV32 and about 0.75 on Ibex. OpenAI is partnering with Synopsys to train GPT-Synopsys to use EDA tools the way expert designers do.

Why it matters: compute is the input to nearly everything in knowledge work, and smaller, faster, lower-power chips lower its cost. If models can routinely shave area and power from the chips that run them, that is a feedback loop on the cost of intelligence, and silicon area translates fairly directly into manufacturing cost.

Caveats: the Jalapeño and AMD figures are OpenAI's own reporting, and the AMD number is an upper bound. The RISC-V cores are microcontroller-class, not flagship designs. Scoring from a specification lets models omit features like performance counters, which may inflate results. OpenAI itself says a more production-like test would require those features, set a tighter area limit and check timing after physical synthesis.

Source: OpenAI