SIAgentNews.com
← All news
DevelopingOpenAIChips

OpenAI models achieve double-digit efficiency gains in chip design

The AI company reported that its models helped reduce power and area for both its own custom hardware and AMD's flagship chips.

· First reported by datacenterdynamics.com

14 percentmaximum PPA improvements on AMD GPU IP

The short version

  • OpenAI used GPT-o3 mini to improve its in-house Jalapeño inference chip.
  • The model reduced standard-cell area in the vector unit by 13 percent.
  • Collaborations with AMD yielded up to 14 percent gains on flagship GPU IP.
  • OpenAI is partnering with Synopsys to create a model that mimics human chip engineering workflows.
  • Testing showed models perform better when designing from scratch than when optimizing existing designs.

Why it matters

This demonstrates that AI can accelerate the physical engineering of hardware, not just software. It suggests a future where chip design cycles are significantly shortened by automated code rewriting.

Optimizing internal hardware

OpenAI recently shared how its large language models are being used to refine the physical architecture of computer chips. The company utilized its models to assist in the development of Jalapeño, which is its own custom inference chip designed for internal use. These models helped engineers navigate the complex trade-offs required to make hardware more efficient.

The specific model used for this task was GPT-o3 mini. During the early 2025 design phase, the AI helped achieve what OpenAI described as non-trivial PPA improvements. PPA stands for power, performance, and area, which are the three primary metrics engineers use to judge how well a chip functions.

The results were measurable and specific to different parts of the hardware. The model successfully reduced the standard-cell area in the vector unit by 13 percent. This helped to make the overall chip more compact and efficient while maintaining its core functionality.

Collaborations with AMD

Beyond its own internal projects, OpenAI worked with AMD to improve third-party hardware designs. This partnership focused on delivering significant gains on the company's flagship GPU IP. A GPU is a specialized processor primarily used for graphics and high-performance computing tasks like training AI.

The AI-driven improvements reached up to 14 percent relative to heavily optimized baselines. These gains were not just theoretical; AMD has reportedly included these specific changes in a recent chip design that is currently in the manufacturing phase. The company plans to add even more changes to the production design files for future versions of the hardware.

OpenAI also observed single-digit PPA percentage gains across key blocks of an upcoming CPU. These CPUs were designed using code written by an OpenAI model. The company noted that the AI identified further improvements by rewriting the SystemVerilog RTL, which is the code that describes the chip's circuits.

Accelerating the design cycle

One of the most significant claims from OpenAI involves the speed at which these AI models can produce results. The company noted that the improvements made by the AI were equivalent to a year's worth of progress achieved in less than five months. This suggests that AI can significantly compress the timeline for hardware development.

To understand these capabilities, OpenAI tested five different models on two open-source 32-bit RISC-V cores. These tests were designed to measure how well the models could handle different types of engineering tasks. The models were asked to either increase the throughput of an existing core or design a new one from scratch based on specific requirements.

The testing revealed a distinct pattern in how the models performed. OpenAI found that all five models scored higher when they were asked to design a core from written requirements rather than trying to improve an existing design. This indicates that the models are currently more effective at creative architecture generation than iterative optimization.

The future of EDA tools

OpenAI is moving toward a more specialized approach to hardware engineering by partnering with Synopsys. This collaboration aims to create a new model called GPT-Synopsys. The goal is to train this model to use Synopsys electronic design automation, or EDA, tools in a way that mimics the actual process used by human engineers.

EDA tools are the software suites that engineers use to design, simulate, and verify integrated circuits. By training a model specifically on these tools, OpenAI hopes to create a more seamless workflow for chip design. This would allow the AI to interact with the same software environments that professionals use every day.

Looking ahead, OpenAI expects to see even more PPA improvements on both GPU and CPU designs relative to the most optimized baselines. The company intends to continue using these models to push the limits of what hardware can do. This transition from general-purpose AI to specialized engineering tools marks a significant step in how the industry approaches silicon manufacturing.

What to watch

  • How the GPT-Synopsys model will perform compared to current human-led EDA workflows.
  • Whether the 14 percent gains on AMD hardware translate to significant consumer performance boosts.
  • If other chip manufacturers will adopt OpenAI's methods for their own custom silicon.

Sources and links

First reported by datacenterdynamics.com. 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