OpenAI’s Models Redesigned AMD’s Flagship GPU — 14% Better

Human Verified

Share to: , and

Computer chip with glowing circuit traces being redesigned by streams of light

OpenAI says its models are now improving the chips that power artificial intelligence — including other companies’ hardware. Its models rewrote the circuits of AMD’s processor, delivering up to 14 percent better power, performance and area, and AMD has put some of the changes into a design sent for manufacturing.

The AMD Result

The headline number is the GPU result. OpenAI says model-proposed rewrites delivered up to 14 percent better power, performance and area on AMD’s flagship graphics IP, against baselines AMD’s own engineers had already heavily optimised. The models rewrote SystemVerilog RTL and checked every change with Synopsys tools.

AMD verified the work on its own side. The company put some of the model-written changes into a recent design sent for manufacturing and wrote the rest into production design files for future versions. AMD also verified single-digit percentage gains across key blocks of its next-generation flagship CPU.

Jalapeño Proved It First

Before AMD, OpenAI tested the approach on its own hardware. Engineers began using GPT-o3 mini to help design Jalapeño, its first custom inference chip, in early 2025. The model cut the vector unit’s standard-cell area by about 13 percent and the matrix unit’s by about 10 percent.

OpenAI is candid about the limits. The production examples are a small set, the area figures are team estimates, and the CPU gains carry no exact number. The from-specification scores may also be flattered, since models could omit features the original cores carried.

How the Models Were Tested

To measure the capability systematically, OpenAI tested five model generations on two open-source 32-bit RISC-V cores, PicoRV32 and Ibex. One task asked models to improve an existing design’s throughput within an area limit; the other asked them to design a core from written requirements.

Across 9,881 scored runs, GPT-6.2 Astra scored highest on both cores at about 112,000 tokens per run. On PicoRV32 it scored roughly twice as well as GPT-6 Astra, the second-best model. On Ibex, its score kept climbing with budget while rivals flattened.

The from-specification task flattered every model: all five scored higher designing a core from requirements than improving an existing one. OpenAI concedes the advantage may be overstated, since models could omit features such as performance counters that the original cores carried.

Our Read

The AMD result is the news, not a benchmark. Changes written by a model are already in a design headed for manufacturing, which is a stronger claim than any leaderboard score. That is why the small sample still matters: one production win beats a thousand simulated ones.

The pattern is now familiar at OpenAI: models improving the work that builds the next models. Its agents already finish half of some research tasks, and its chips are now partly designed by its own models. The loop is closing — and the caveats are the honest part.

Why It Matters

If models can find a year’s worth of optimisation in under five months, chip design cycles shorten — and so does the cost of the hardware that runs AI. The bottleneck shifts from engineering time to fabrication capacity — a different race, and one the industry is in.

The deal also tightens OpenAI’s bond with AMD beyond buying graphics cards. OpenAI already holds rights to buy up to 160 million AMD shares under its compute agreement, and its models now shape the silicon those shares represent. The first planned deployment uses AMD’s Instinct MI450 series GPUs.

Background

OpenAI and Synopsys announced their partnership in September: a multi-year deal to train GPT-Synopsys, a model that uses Synopsys design tools the way expert engineers do. The aim is to put frontier models in the hands of every chip designer, not only the labs that build them.

Source: OpenAI’s report

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.