Skip to content
AI.info

The Pulse

OpenAI says its models improved AMD GPU designs by up to 14%

OpenAI reports that model-written circuit changes improved power, performance and area on AMD’s flagship GPU intellectual property. The company also detailed changes to its Jalapeño chip and tests on two open-source processor cores.

OpenAI says its models improved AMD GPU designs by up to 14%

OpenAI’s models produced up to a 14% improvement in power, performance and area relative to heavily optimized baselines on AMD’s flagship GPU intellectual property, the AI company says. In a November 18 post, OpenAI detailed the model-written circuit changes and described related work on its own Jalapeño inference chip.

Some changes have been included in an AMD design sent for manufacturing, while others have been added to production design files for future versions. The post also described tests of model-designed versions of two open-source processor cores.

Changes to AMD circuits

OpenAI’s model rewrote SystemVerilog RTL, the code that describes a chip’s circuits, to find further improvements. The model used Synopsys tools to check that each rewritten block retained its behavior and to assess trade-offs in power, performance and area.

OpenAI characterized the GPU work as roughly a year of progress achieved in less than five months. AMD also verified single-digit percentage gains in power, performance and area across key blocks of its upcoming next-generation flagship CPU using model-written designs, according to the post.

Jalapeño design changes

OpenAI began using GPT-o3 mini to help design Jalapeño in early 2025. The team estimates that model-proposed changes reduced standard-cell area by about 13% in the vector unit, which handles parallel computing tasks, and 10% in the matrix unit, an arithmetic component.

The company says the optimizations were proven during silicon bring-up. The changes included alternative logic implementations, different ways to divide computations across pipeline stages, and opportunities to share hardware and reduce intermediate storage.

One matrix-unit multiplier redesign used about 15% less cell area in synthesis while reducing latency from two clock cycles to one, OpenAI reported. A separate adder redesign reduced cell area by about 15% while keeping its two-cycle latency; further changes to dot-product logic and pipeline organization contributed additional savings across the block.

Tests on open-source cores

The company also tested successive model generations on two open-source 32-bit RISC-V cores: PicoRV32 and Ibex. In one task, models received the original hardware description and were asked to increase throughput, measured as an estimate of instructions executed per second.

Designs received a zero if they failed a simulation check or used more than 16 times the original core’s area. Passing designs were measured on workloads hidden from the models, with throughput estimates based on instructions per clock cycle and a maximum clock frequency estimated using Yosys.

At a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores, OpenAI said. For PicoRV32 at that budget, its score was about twice that of GPT-6 Astra.

In a second task, models designed cores from written requirements and templates showing how each core connected to the rest of the system. At about 112,000 tokens per run, all five models scored higher on both cores in that task than when improving the existing designs. The evaluation included 9,881 scored runs across the two tasks, both cores and all token budgets.

The company cautioned that starting from written requirements allowed models to omit features in the original cores, including performance counters and additional instructions, which could explain part of the advantage. It said requiring those features, imposing a tighter area limit and checking timing after physical synthesis would make the test more representative of production chip design.

Work with Synopsys

OpenAI is also working with Synopsys to train GPT-Synopsys, a new model intended to use the company’s electronic design automation tools as expert chip designers do. The tools are used by engineers to design and verify chips.

Source

  1. OpenAI

Companies in this story

Related stories

Explore

More articles