OpenAI says its models improved AMD graphics chip design by up to 14%


OpenAI also says AMD verified unspecified single-digit gains in key blocks of its next flagship CPU, with some model-generated changes sent for manufacturing.
OpenAI says its superintelligence (SI) models improved circuit designs for AMD’s flagship graphics processing unit, or GPU, by up to 14% in power, performance and area compared with heavily optimized baselines. The figure is a best case rather than an average.
In its November 18 disclosure, OpenAI also said AMD had verified unspecified single-digit percentage gains on those measures in key blocks of its next flagship central processing unit, or CPU. OpenAI has not identified the GPU blocks or their share of the chip, and no independent party has checked the GPU and CPU results, Thorsten Meyer AI said.
The companies’ collaboration predates the design work: in October 2025, they announced a separate multiyear agreement for OpenAI to deploy 6 gigawatts of AMD Instinct GPUs for its SI infrastructure.
Earlier: OpenAI says its models optimized chip circuits and core designs
OpenAI said its models proposed changes by rewriting circuit descriptions in SystemVerilog register-transfer level, or RTL, code. RTL describes how data moves between storage registers and how circuit logic processes it. Design tools then translate that description into a gate-level design before the physical chip is laid out and manufactured.
OpenAI said Synopsys tools checked whether each rewrite preserved the original design’s behavior and assessed the trade-offs in power, performance and area. AMD engineers decided which changes entered the design, Thorsten Meyer AI said.

The company says its work with GPT-6 Astra improved circuits used in its Jalapeño chip and separate RISC-V processor designs, extending its use of models in chip design.
OpenAI described the work as roughly a year of conventional chip-design progress achieved in under 5 months, Tom’s Hardware said. OpenAI expects further improvements in both designs from the revised baseline.
OpenAI previously said model-proposed changes reduced standard-cell area by about 13% in the warp engine and 10% in the matrix unit of Jalapeño, its inference chip developed with Broadcom. It said the changes were validated when the first chips came back. It also tested models on PicoRV32 and Ibex, two public processor cores; GPT-6.2 Astra achieved the highest average score across both in 9,881 scored runs. Unlike those public cores, AMD’s designs are proprietary.
OpenAI said AMD included some model-generated changes in a design sent for manufacturing and placed others in production design files for future versions. No product or release date has been identified. Testing finished chips would be needed to establish whether the design-stage gains carry through to power use or speed.
OpenAI and Synopsys have also announced a multiyear agreement to develop GPT-Synopsys, a specialized model intended to use Synopsys chip-design tools. The companies plan to make it available to customers worldwide.
2 min