OpenAI said Wednesday that AMD included some model-proposed circuit changes in a design already sent for manufacturing. The models rewrote circuit code, delivering up to 14% improvement on power, performance and area measures against heavily optimized baselines for AMD’s flagship graphics processor design. OpenAI disclosed no breakdown of the three measures, no named AMD product and no performance numbers from finished AMD silicon.

The companies already have a GPU purchase agreement announced in October 2025. The first planned deployment under that agreement was to use AMD Instinct MI450 series GPUs. The November 18 design disclosure names no chip and provides no link to that deployment.

What Changed

AI-generated summary, reviewed by an editor. More on our AI guidelines.

Circuit changes

The models rewrote SystemVerilog RTL, code describing how circuits operate. Synopsys software checked whether each rewrite preserved the block’s behavior and evaluated its power, performance and area trade-offs. Those checks do not establish the performance of a finished product. AMD also put other changes into production design files for future versions.

OpenAI separately said AMD verified single-digit percentage gains on those measures in blocks of its next-generation flagship CPU.

OpenAI’s own Jalapeño chip, whose working silicon was described in August, provides a separate hardware example. In Wednesday’s account, OpenAI estimated model-proposed changes across the broader design effort reduced standard-cell area by about 13% in the vector unit and 10% in the matrix unit. OpenAI said the changes were validated during silicon bring-up.

Simulation boundaries

The new benchmark tested PicoRV32 and Ibex, embedded RISC-V processor cores. Its 9,881 evaluated runs covered both tasks, both cores and all token budgets. Designs had to pass simulation checks and stay within an area ceiling of 16 times the original core. For designs that passed, estimated throughput multiplied the geometric mean of instructions per clock cycle across hidden workloads by an estimated maximum clock frequency. Scores set each core’s best observed throughput under those constraints to 1, with zero for failed checks or exceeding the area ceiling.

At a comparable budget of roughly 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores in OpenAI’s circuit rewrite benchmark.

At that same budget, all five tested models scored higher when building from written requirements than when modifying existing code. OpenAI’s description of the benchmark said these designs could omit features such as performance counters or additional instructions. OpenAI identified tighter area limits and timing checks after physical synthesis as steps toward a more representative production test.

A separate CPU hardware trial

A separate October 29 preprint by University of Tokyo researchers Leo Itoh and Junichiro Kadomoto compared a short diagnostic with longer tests of a modified CPU.

Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.

In that study, a modified CPU improved in a 10-iteration diagnostic but used 9.98% more cycles than the preceding design over 8,000 CoreMark iterations on a reprogrammable chip. The team rejected that candidate after repeated board runs. Longer simulation reproduced the slowdown. Tracing the circuit exposed a predictor state that left execution persistently slower.

The study did not directly test OpenAI’s AMD claims or its RISC-V figures.

A planned design service

Under their September 30 agreement, Synopsys and OpenAI plan GPT-Synopsys, a specialized model operating chip-design tools on OpenAI-hosted infrastructure. The planned service bundles compute, model access and software licenses. Early customer engagements were already underway at the announcement. The companies said customer design data in the planned service will be encrypted and excluded from model training, with configurable retention and access controls.

OpenAI president Greg Brockman described the work then as “helping engineers explore more designs and get to a working chip faster.”

Frequently Asked Questions

What did OpenAI report about AMD’s GPU design?

OpenAI reported up to 14% improvement on power, performance and area measures against previously optimized designs for AMD’s flagship GPU intellectual property.

What has AMD sent for manufacturing?

OpenAI says AMD submitted a design that incorporated some model-proposed circuit changes. OpenAI did not identify the product or provide measurements of finished AMD silicon.

What did OpenAI measure on Jalapeño?

OpenAI estimated a roughly 13% reduction in standard-cell area in the vector unit and 10% in the matrix unit. The company says the changes were validated during silicon bring-up.

Which processor cores did OpenAI test?

OpenAI used two embedded RISC-V cores, PicoRV32 and Ibex. GPT-6.2 Astra had the highest average score on both in the circuit-rewrite task at about 112,000 tokens per run.

What is GPT-Synopsys?

It is a planned model from Synopsys and OpenAI designed to operate chip-design tools. The planned service combines compute, model access and software licenses.

AI-generated summary, reviewed by an editor. More on our AI guidelines.

AMD Offered Meta and OpenAI 20% of Stock to Sell AI Chips
AMD offered warrants for 20% of its stock to Meta and OpenAI in four months. The AI chip market's circular financing raises serious questions.
GPT-6.2 Astra and Sol: What OpenAI's New Models Can Do
OpenAI says GPT-6.2 Astra and Sol finish professional, legal and scientific work for a fraction of the old cost per task. Independent testers confirm part of that, and OpenAI's own system card flags a regression it calls concerning.
Sam Altman Says PC Memory Is Better Used in Data Centers
OpenAI's chief says the memory in home computers would do more good in the cloud. The idle-time figure is his own, and data centers running closed models are his company's main business.
AI News

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai