Latest
AI Affairs, home

Thursday 19 November 2026

Technology

OpenAI says its models improved AMD chip designs in under five months

The company says AMD has sent some model-proposed GPU changes for manufacturing and verified smaller gains in key blocks of an upcoming CPU.

OpenAI's headquarters at 1515 Third Street, with palm trees and overhead wires
Photo: Coolcaesar, CC BY 4.0, via Wikimedia Commons (cropped)

OpenAI said in a research article published on 18 November 2026 that its models had improved parts of AMD’s flagship GPU design by up to 14% against heavily optimised baselines in less than five months. Some of the model-proposed changes have gone into a design AMD sent for manufacturing, according to OpenAI. The work extends a collaboration between the companies from supplying computing capacity to changing the circuits inside future chips.

Key points

  • OpenAI reports improvements of up to 14% against heavily optimised baselines on AMD’s flagship GPU IP, and says AMD has sent some changes for manufacturing.
  • OpenAI says AMD verified single-digit percentage gains in key blocks of an upcoming CPU.
  • GPT-6.2 Astra led OpenAI’s tests on two open-source processor cores at a comparable budget of about 112,000 tokens per run.
  • OpenAI is working with Synopsys to train a model to use chip-design and verification tools.

AMD’s GPU design reaches manufacturing

The GPU work began with SystemVerilog RTL, code that describes how a chip’s circuits behave. OpenAI says its model rewrote circuits responsible for calculations, control and moving data, then used Synopsys tools to check whether each rewrite preserved the original behaviour and to assess power, performance and area. Those three measures can pull in different directions: a circuit that finishes an operation in fewer clock cycles may use more space, or its cycles may take longer.

OpenAI describes its gain on AMD’s flagship GPU IP as an improvement of up to 14% relative to heavily optimised baselines. It characterises the work as roughly a year of progress achieved in less than five months. According to OpenAI, AMD put some changes into a recent design sent for manufacturing and placed others in production design files for future versions. OpenAI also says AMD verified single-digit percentage improvements in power, performance and area across key blocks of its upcoming next-generation flagship CPU using model-written designs.

The companies have a separate commercial relationship. Under a deal announced in October 2025, OpenAI committed to 6 gigawatts of AMD GPU capacity and secured warrants to buy up to 160 million AMD shares, Crypto Briefing reported. The warrants vest against deployment milestones, according to the publication. AI Affairs has reported AMD’s quarterly charge on its OpenAI and Meta chip warrants.

Jalapeño’s vector and matrix units shrink

OpenAI began using GPT-4o in early 2025 to help design Jalapeño, its first custom inference chip. The company estimates that model-proposed changes reduced standard-cell area — the space occupied by the basic circuit components used to build a chip — by about 13% in its vector unit and 10% in its matrix unit. OpenAI says the optimisations were proven when the manufactured chip was brought into operation.

In the matrix unit, OpenAI says one redesigned multiplier used about 15% less cell area when its circuit was synthesised and completed its operation in one clock cycle rather than two. A separate adder redesign reduced cell area by about 15% without changing its two-cycle latency. The company attributes the broader savings to changes across arithmetic circuits and to how work was organised through successive stages of the chip.

Multiplying two numbers could take one clock cycle rather than two if that redesigned multiplier performed the operation.

GPT-6.2 Astra leads two RISC-V tests

OpenAI also tested successive models on PicoRV32 and Ibex, two open-source 32-bit RISC-V processor cores intended for embedded workloads. In one task, each model received an existing hardware description and tried to increase throughput, meaning the number of instructions completed per second. In another, models built a core from written requirements and templates showing how it connected to the rest of a system. The full evaluation comprised 9,881 scored runs across both tasks, both cores and the tested token budgets.

OpenAI checked proposed designs in simulation, assigning a score of zero to those that failed a correctness check or exceeded 16 times the original core’s area. For designs that passed, it measured instructions per clock cycle on workloads hidden from the models, combined those results across workloads and multiplied by a maximum clock frequency estimated using Yosys. At a comparable budget of about 112,000 tokens per run, OpenAI reports that GPT-6.2 Astra had the highest average score on both cores.

At that budget, all five models scored higher when designing from written requirements than when modifying an existing core. The company says the written-requirements task allowed models to omit features present in the original designs, including performance counters and additional instructions, which could account for part of the advantage. It says requiring those features, limiting area more tightly and checking timing after physical synthesis would make the evaluation more representative of production chip design.

OpenAI and Synopsys train GPT-Synopsys

OpenAI says it is partnering with Synopsys to train GPT-Synopsys to use electronic design automation tools, the software chip designers use to develop and verify circuits. The partnership follows OpenAI’s use of Synopsys tools to check the behaviour and power, performance and area of its proposed AMD circuit rewrites.

Sources

Topics: Chips, Foundation models

What should we cover? Tell us

The weekly briefing

The AI Affairs weekly briefing: the week's 10 most important AI stories in business, government and markets, every Monday. We'll email you a link to confirm. Every email has a one-click unsubscribe. Privacy policy