Money and business in the Middle East.

AI

AMD sends chip design with OpenAI-proposed changes to manufacturing

OpenAI said its models delivered up to 14% design gains on AMD's flagship GPU IP

· Last verified: 19 Nov 2026 (OpenAI)

Summary

  • OpenAI said its models delivered up to 14% design gains on AMD's flagship GPU IP
  • AMD included some model-proposed changes in a design already sent for manufacturing, OpenAI said
  • Expert-level circuit rewriting moves AI from assisting chip designers to shipping production silicon

The latest

Up to 14% power, performance and area improvement on AMD's flagship GPU IP, measured against heavily optimized baselines, came from an OpenAI model rewriting the SystemVerilog code describing the chip's circuits, the company said in a research post dated November 18, 2026. AMD has already shipped a recent design containing some of those changes to manufacturing, and filed others into its production design files for future versions.

Details

  • The AMD result: OpenAI said the 14% gains on AMD's flagship GPU IP came against baselines its engineers describe as already heavily optimized, and characterized the work as roughly a year of progress compressed into under 5 months. It expects further PPA improvements on both GPU and CPU.
  • CPU side: OpenAI said AMD independently verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU, using designs written by the model. Neither the shipping timeline for that CPU nor the specific blocks involved was disclosed.
  • The company's own chip: The approach began with Jalapeño, OpenAI's first custom inference chip. Engineers started using GPT-4o on the design in early 2025, and the team estimates model-proposed changes cut standard-cell area by about 13% in the warp engine and 10% in the matrix unit, optimizations OpenAI said were silicon proven during bring-up.
  • Circuit-level specifics: In the matrix unit, one multiplier redesign used about 15% less cell area in synthesis while cutting latency from two clock cycles to one, according to OpenAI. A separate adder redesign saved roughly 15% of cell area while keeping its two-cycle latency, with further savings from dot-product logic and pipeline reorganization.
  • How it works: OpenAI said the model reasons about register-transfer-level code at the level of an expert microarchitect, rewriting circuits that perform calculations, control operations or move data. It used Synopsys tools to verify each rewrite preserved the block's behavior and to weigh power, performance and area trade-offs.
  • The benchmark: To track generational progress, OpenAI tested successive models on two open-source 32-bit RISC-V cores: PicoRV32, built for FPGAs, and Ibex, the embedded processor inside the OpenTitan security chip. Each model was given the original hardware description and asked to raise throughput.
  • Model scores: At a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra posted the highest average score on both cores, and GPT-6.1 Sol outscored GPT-5.6 Sol on both, OpenAI said. On PicoRV32 at that budget, GPT-6.2 Astra scored roughly twice GPT-6 Astra.
  • Spec beats repair: In a second task, models were handed written requirements and connection templates rather than existing code. At about 112,000 tokens, all five models scored higher designing a core from scratch than improving the existing one, on both cores. The full evaluation covered 9,881 scored runs.
  • The Synopsys deal: OpenAI said it is partnering with Synopsys to train GPT-Synopsys, a model built to operate electronic design automation tools the way expert chip designers do. The stated aim is grounding model development in trusted evaluators and widening the range of design problems addressed. No timeline was given.

Background

PPA — power, performance, area — is the standard yardstick for chip design quality. RTL is the code layer where engineers describe circuit behavior before synthesis turns it into physical layout. Gains against an already optimized RTL baseline are historically slow and incremental.

Between the lines

The manufacturing detail is what separates this from a benchmark release. Model scores on open-source RISC-V cores measure capability in a sandbox; a design sent to a fab carries commercial risk AMD chose to accept. Note the pattern in OpenAI's own evaluation: models scored higher writing cores from written requirements than improving existing ones, suggesting the harder task remains optimizing work humans already tuned.

What's next

Watch for AMD's next-generation flagship GPU and CPU launches and whether either company quantifies the shipped gains, plus any release detail on GPT-Synopsys.

Source: OpenAI

AMD sends chip design with OpenAI-proposed changes to manufacturing · INXEN