Skip to content
D-CSIL™

AI News · 2026-11-19 · 7:03 AM CT

OpenAI says AI-written changes reached an AMD GPU design sent for manufacturing

Paul Derrington, Jr.

By Paul Derrington, Jr.

CEO | Chief Engineer | Research Engineer | Editor, D-CSIL

4 min read

Follow @PaulDerrington on X

Key Takeaways

OpenAI reports up to 14% improvement in power, performance and area (PPA) for selected blocks of AMD's flagship GPU design, not a 14% faster finished GPU. It says AMD included some changes in a design sent for manufacturing, not that a resulting product has shipped. These results have not been independently verified here. OpenAI has an AMD GPU purchasing relationship and a conditional warrant for AMD shares, plus a revenue-sharing partnership with Synopsys.

Die shot of an AMD Zen 4 Raphael processor, showing the core complex die of a Ryzen 7 7800X3D
Photo: Fritzchens Fritz / Wikimedia Commons (CC0)

Early 2025: the method starts on OpenAI's own chip

OpenAI started using GPT-o3 mini to help design Jalapeño, its first custom inference chip, in early 2025. In August 2026 the company reported the first results: AI help took Jalapeño from initial design to tapeout — the handoff of design data to manufacturing — in nine months.

The November 18 post fills in where the gains came from. In the vector unit, which runs parallel computing tasks, the savings came from floating-point adders and multipliers, conversion functions, reduction logic and the special-function unit. Across the design, the team estimates model-proposed changes reduced standard-cell area — the space taken by the chip's basic logic circuits — by about 13% in the vector unit and 10% in the matrix unit. One multiplier redesign used about 15% less cell area in synthesis while cutting latency from two clock cycles to one. OpenAI says these optimizations were silicon-proven during Jalapeño's bring-up: measured on the real chip, not just in simulation.

November 18: the same rewrites land in AMD production designs

OpenAI says that, working with AMD, its model rewrote parts of the flagship GPU's SystemVerilog RTL — the code that describes the chip's circuits before physical layout. For each rewrite, the model used Synopsys design software to check that the changed block still behaved as before and to measure the trade-offs. OpenAI reports up to 14% improvement for selected design blocks against heavily optimized baselines in PPA — power, performance and area, the three measures chip designers balance. OpenAI calls that roughly a year of design progress achieved in less than five months.

OpenAI says AMD has already included some of them in a recent design it sent for manufacturing, and added others to its production design files for use in future versions. Separately, OpenAI says AMD verified single-digit percentage gains in PPA across key blocks of its upcoming next-generation flagship CPU, using designs written by the model. Sending a design for manufacturing is not the same as shipping a finished GPU to customers.

The benchmark behind it — and its built-in caveat

To measure how chip-design ability changes from one model generation to the next, OpenAI ran 9,881 scored evaluations on two open-source 32-bit RISC-V cores: PicoRV32, a compact core for reprogrammable FPGA chips, and Ibex, the embedded core inside the OpenTitan security chip. At a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores; on the existing PicoRV32 design, its score was about twice GPT-6 Astra's.

OpenAI states the caveat itself. Every model tested scored higher when designing a core from a written specification than when improving the existing design — partly, OpenAI notes, because a from-spec design can leave out features such as performance counters that the test workloads do not reward. Requiring those features, setting a tighter limit on circuit area, and checking timing after physical synthesis would make the test more representative of production chip design.

What the figures establish, and the commercial relationships

The 14% is an 'up to' PPA figure from OpenAI for GPU design work, not a measured 14% increase in finished-GPU speed. PPA covers power, performance and area; the post does not define a combined percentage formula or provide a per-metric breakdown for named AMD blocks. These results have not been independently verified here. The cited post does not provide measurements from finished AMD silicon carrying these changes.

OpenAI has a financial interest in this work. AMD's August 2026 filing describes its GPU deployment agreement with OpenAI and a warrant allowing OpenAI to buy up to 160 million AMD shares, subject to purchase milestones, stock-price or performance thresholds, and other conditions. That is a conditional right, not evidence that OpenAI already owns those shares.

GPT-Synopsys is an existing development partnership, announced on September 30, rather than a new November launch. Synopsys says the companies agreed to develop the chip-design model, license its design tools and share revenue, with early customer engagements underway. The cited announcements do not specify a general release date.

Disclaimer: Derrington Collaborative AI is my personal business, operated independently of and outside my official U.S. Government position and duties. The views and opinions expressed on this site and on any and all of my social media accounts are my own. They do not reflect the official policy or position of the Department of the Army, the Department of Defense, or the U.S. Government. Mention of any company, product, or service does not constitute an official endorsement.