SDSignal Desk

OpenAI says its models now shave area off production chips, including AMD’s next flagships

· Updated · Signal Desk newsroom

Analysis of reporting by OpenAI. How we work

Image: OpenAI

OpenAI reports model-proposed changes cut about 10–13% of standard-cell area in its Jalapeño chip and found up to 14% PPA gains on AMD GPU IP. The AI-designs-AI-hardware loop is real, and the evidence is still mostly in-house.

Key takeaways

  • OpenAI estimates model-proposed changes cut standard-cell area about 13% in Jalapeño’s vector unit and 10% in its matrix unit, silicon-proven at bring-up.
  • One multiplier redesign used about 15% less area and cut latency from two cycles to one.
  • With AMD, OpenAI reports up to 14% PPA improvement on flagship GPU IP versus heavily optimized baselines; some changes are already in a design sent for manufacturing.
  • AMD verified single-digit percentage PPA gains on key blocks of its upcoming flagship CPU.
  • On open RISC-V benchmarks, GPT-6.2 Astra scored highest; OpenAI is training GPT-Synopsys with Synopsys to use EDA tools like an expert.

From chatbots to circuits

OpenAI’s chip-design post is a progress report on one of the most consequential uses of frontier models: designing the hardware that runs them. In August the company said its models helped build Jalapeño, its first custom inference chip. Now it gives specifics. Starting with GPT-o3 mini in early 2025, models proposed changes to floating-point adders, multipliers, conversion functions, reduction logic and dot-product circuits. OpenAI estimates those changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit, and says they were silicon-proven during bring-up.

Individual results are concrete. One multiplier redesign used about 15% less cell area in synthesis and cut latency from two clock cycles to one, which is a rare win on both size and speed. A separate adder redesign saved about 15% area at the same latency. These are the kinds of micro-architectural gains human teams chase for months.

The AMD result is the headline

The bigger claim involves someone else’s silicon. Working with AMD, OpenAI reports up to 14% improvement in power, performance and area (PPA) against heavily optimized baselines on flagship GPU IP, by rewriting SystemVerilog RTL. OpenAI calls this roughly a year of progress in under five months. AMD has put some changes into a design sent for manufacturing and others into production files for future versions. AMD also verified single-digit percentage PPA gains on key blocks of its next flagship CPU using model-written designs. Each rewrite was checked with Synopsys tools to confirm unchanged behavior.

That third-party adoption matters more than any benchmark. A chipmaker putting model-written RTL into a tape-out is a strong vote of confidence. There are caveats. “Up to 14%” is a best case, not an average. The comparison of a year of progress in five months is OpenAI’s framing. And we haven’t seen AMD’s own account of how much human review each change needed. Still, this is a meaningful step past demos.

Benchmarks, and an honest limitation

To track progress between model generations, OpenAI tested models on two open-source RISC-V cores, PicoRV32 and Ibex, both by optimizing existing RTL and by designing from a written specification. Across 9,881 scored runs, GPT-6.2 Astra scored highest at a budget of about 112,000 tokens, and on PicoRV32 its score was about twice GPT-6 Astra’s. All models did better designing from a specification than improving existing code. OpenAI notes why that may flatter the results: designing from scratch let models skip features like performance counters that don’t help throughput. It suggests adding those requirements, tighter area limits and post-synthesis timing checks to make the test more realistic. That kind of self-critique is welcome.

The larger story is the loop. Models help design chips; better chips train better models. OpenAI’s own research-acceleration reporting frames recursive self-improvement as something to monitor, and hardware is part of that loop. Partnering with Synopsys on GPT-Synopsys, a model trained to use EDA tools like an expert designer, makes the trend industrial. It also raises a competitive question. If frontier labs control the best chip-design models, chipmakers become customers of the labs they supply.

Who it matters to

Chip designers
Model-proposed RTL rewrites are moving into production flows; verification skills become more central.
AMD and rival chipmakers
Early access to model-driven PPA gains could be a competitive edge, and a dependency on OpenAI.
EDA vendors
Synopsys is positioning as the trusted verifier inside AI design loops.
AI safety watchers
AI improving AI hardware is a concrete channel for self-acceleration.

What to watch

  1. AMD’s own disclosure of how model-written changes performed in shipping silicon.
  2. Release details and availability for GPT-Synopsys.
  3. Tougher RISC-V tests with feature requirements and post-synthesis timing.
  4. Whether other chipmakers, or rival labs, announce similar results.
  5. Performance of Jalapeño in OpenAI’s production inference fleet.

FAQ

What is Jalapeño?

OpenAI’s first custom inference chip, whose design used model-proposed improvements starting in early 2025.

What did OpenAI do with AMD?

Its models rewrote RTL on AMD’s flagship GPU IP for up to 14% PPA improvement; some changes are in a design sent for manufacturing.

Which model performed best on RISC-V tests?

GPT-6.2 Astra had the highest average score on both PicoRV32 and Ibex at a budget of about 112,000 tokens.

What is GPT-Synopsys?

A model OpenAI is training with Synopsys to use electronic design automation tools the way expert chip designers do.

Read the original

Continue at the source.

OpenAI

Companies: OpenAI

More on OpenAI

All OpenAI coverage →