SIAgentNews.com
← All news
DevelopingOpenAIChips

OpenAI helps AMD improve chip designs by 14% using AI models

OpenAI's AI models helped AMD optimize its upcoming flagship GPU and CPU designs, potentially condensing a year of engineering work into five months.

· First reported by Tom's Hardware

14%PPA improvement for GPU circuitry

The short version

  • OpenAI's models achieved a 14% PPA (power, performance, and area) improvement for AMD's upcoming flagship GPU.
  • The AI-driven optimization condensed a year's worth of conventional chip design work into under 5 months.
  • The models were able to identify improvements by rewriting SystemVerilog RTL, the standard language for high-end chip design.
  • OpenAI is also working with Synopsys to develop a specialized model for chip design tools called GPT-Synopsis.

Why it matters

This demonstrates that AI can perform high-level engineering tasks, such as microarchitecture optimization, much faster than human teams. It signals a shift where AI will be used to accelerate the production of the very hardware required to run AI.

Accelerating hardware design

OpenAI recently shared details regarding a partnership with AMD to optimize the hardware maker's next-generation chips. Engineers used undisclosed AI models to refine the designs of AMD's upcoming flagship GPU and CPU. These models were able to identify specific improvements by rewriting the code that describes the physical circuits of the chips.

The results of this collaboration were significant for the hardware's PPA. PPA stands for power, performance, and area, which are the three primary metrics used to judge the efficiency of a semiconductor design. The AI helped achieve a 14% improvement in these metrics for the GPU circuitry, while providing single-digit gains for the processor.

The speed of this development was particularly notable for the industry. OpenAI claims that the AI's optimization efforts were equivalent to condensing a year's worth of traditional chip design work into under 5 months. This represents a time crunch that the company describes as unprecedented in the field of hardware engineering.

How the AI reasons on circuits

To achieve these results, the AI models worked directly with SystemVerilog RTL. This is the standard language used by engineers to describe the logic and structure of high-end integrated circuits. The models were able to reason on this code at a level comparable to an expert microarchitect, which is the person responsible for the high-level design of a processor.

OpenAI tested the capabilities of its high-end models by asking them to improve specific chip cores, such as the PicoRV32 and Ibex. The models were given 112,000 tokens per run and instructed to optimize for throughput. Throughput is a measure of how much work a chip can do, calculated by combining instructions-per-clock with the duration of those clock cycles.

The researchers used a scoring system to evaluate the AI's output. A design received a score of 1.0 for the best results under specific constraints, while a score of zero was given if the chip failed to work or exceeded the previous area by more than 16x. Interestingly, the models performed better when they were asked to redesign the chips from scratch based on initial specifications rather than trying to optimize existing code.

Broader AI hardware projects

This partnership is part of a larger trend of OpenAI involving itself in hardware manufacturing. The company previously worked with Broadcom to design a chip named Jalapeño, which is intended specifically for AI inference. Inference is the process where a trained AI model makes predictions or generates content based on new input.

Jalapeño is expected to have small initial deployments around late 2026, with a broader rollout planned for 2027. Early benchmarks for the chip have already shown significant savings in power draw and latency when compared to Nvidia's Blackwell GPUs. These savings are particularly relevant for single-token prediction workloads, which are the tasks where an AI generates text one word at a time.

OpenAI is also looking toward specialized tools for the industry. The company is currently working with Synopsys to train a model called GPT-Synopsis. This model is specifically designed to help engineers use the complex software tools required to design modern semiconductors.

The future of chip design

The success of the AMD partnership suggests that AI will become a standard part of the semiconductor lifecycle. By automating the tedious and complex task of circuit optimization, companies can bring new products to market much faster. This helps bridge the gap between the rapid pace of AI software development and the slower pace of hardware manufacturing.

OpenAI expects to find even more PPA improvements in both the GPU and CPU designs beyond the current baseline. As the models become more adept at understanding microarchitecture, the ceiling for how much they can squeeze out of a design may continue to rise. This could lead to more efficient chips that require less electricity and less physical space.

While the specific models used for the AMD project remain undisclosed, the results prove that high-end LLMs can handle specialized engineering tasks. The industry will likely see more collaborations where AI is used to verify, test, and optimize the very silicon that powers the next generation of computing. This creates a feedback loop where AI helps build better hardware to run even more advanced AI.

What to watch

  • Whether GPT-Synopsis becomes a standard tool for semiconductor companies.
  • The performance of the Jalapeño chip once it hits the market in 2026.
  • How much further the 14% PPA improvement can be pushed by subsequent AI iterations.

Try it

Look up the term 'SystemVerilog RTL' to see the language engineers use to describe how a chip's circuits are laid out.

Sources and links

First reported by Tom's Hardware. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news