ToolNavs Find Useful AI Tools
Submit Sign in
Back to AI information
OpenAI Models Are Now Designing Chips: Jalapeño Saves 13% Area, AMD Gains Nearly a Year

OpenAI Models Are Now Designing Chips: Jalapeño Saves 13% Area, AMD Gains Nearly a Year

AI information • Admin • • 6 views

On November 18, 2026, OpenAI disclosed in an official blog post how its models are contributing to chip design: in its first custom inference chip, Jalapeño, model-proposed changes cut standard-cell area by about 13% in the vector unit and about 10% in the matrix unit — and the optimizations were proven in silicon. This is not a paper benchmark; the model changes made it into a taped-out chip.

Where Jalapeño's Savings Came From

Jalapeño is OpenAI's first custom inference chip, and models have been helping design it since early 2025, starting with GPT-4o. OpenAI broke down exactly where the area savings came from: in the vector unit, floating-point adders and multipliers, conversion functions, and reduction logic were all reworked, with the model trying alternative logic implementations and pipeline splits so hardware could be shared and less intermediate data stored; in the matrix unit, one multiplier redesign used about 15% less cell area in synthesis while cutting latency from two clock cycles to one, and a separate adder redesign saved about 15% area with latency unchanged. Together, those changes produced the 13% and 10% figures.

AMD's Line: Nearly a Year of Progress in Under Five Months

Beyond its own chip, OpenAI applied the same approach to AMD's flagship GPU. The model rewrote SystemVerilog RTL describing the circuits, squeezing up to 14% in combined power, performance, and area gains over an already heavily optimized baseline — roughly a year of progress achieved in less than five months, in OpenAI's words. AMD has already included some of these changes in a design sent for manufacturing, and added others to production design files for future versions. On key blocks of its upcoming flagship CPU, model-written designs also delivered single-digit percentage gains.

What matters here is not any single percentage but the verification: every rewrite had to pass Synopsys EDA tools to confirm behavior was preserved, then be evaluated on power, performance, and area trade-offs. The model is not sketching circuits on paper; it is closing the loop on real production designs with the industry's existing toolchain. OpenAI also published cross-generation evaluations: testing successive models on two open-source 32-bit RISC-V cores, PicoRV32 and Ibex, GPT-6.2 Astra scored highest on both at a budget of about 112,000 tokens per run, and GPT-6.1 Sol beat GPT-5.6 Sol on both — evidence that chip-design ability is genuinely improving with each model generation.

Two things to watch next: whether GPT-Synopsys, the model OpenAI is training with Synopsys, turns this capability into a daily tool for chip engineers, and whether measured data from AMD's next-generation GPU and CPU confirms these percentages. Ordinary readers need not wait for chips to ship to care — once AI lowers the cost of designing chips, the barrier to custom silicon falls with it.

Recommended Tools

More