OpenAI models cut Jalapeño block area by 10–13% and improved AMD GPUs by up to 14%
OpenAI tasked its models with rewriting chip circuit code. Think of it as refactoring, but for a processor. In Jalapeño, cell area fell 13% in the vector block and 10% in the matrix block, and in an AMD GPU the PPA (power, performance, area) gain reached 14%. Synopsys tools verified every edit, and some of the changes have already gone into production. Takeaway: AI is strong where results can be checked automatically.
AI-processed from OpenAI Blog; edited by Hamidun News
OpenAI has described how its models are already improving real chips. Think of refactoring: you take a working function and rewrite it to run faster and use less memory, while it behaves exactly the same. This is the same idea, except that instead of a function, the target is a processor block described in SystemVerilog code (known as RTL).
Results are measured on three parameters at once: power, performance and die area, which the industry shortens to PPA. In Jalapeño, OpenAI's own inference accelerator, the models cut standard-cell area by about 13% in the vector block and by 10% in the matrix block. In an AMD GPU that was already heavily optimized, the PPA gain was up to 14%.
For developers, the takeaway is simple: when a task has automatic verification and a metric, a model can deliver expert-level results.
What the models did in Jalapeño
OpenAI began bringing GPT‑4o into Jalapeño development in early 2025. Even that model found non-trivial improvements.
The vector block handles parallel computation. The savings there came from floating-point adders and multipliers, conversion functions, reduction logic and the special-function unit. The model tried alternative logic implementations and split computations across pipeline stages differently. That let it find places where different operations can share the same hardware and where less intermediate data needs to be stored.
In the matrix block, one multiplier took about 15% less area after rework at synthesis, and its latency dropped from two cycles to one. Another adder became about 15% smaller, and its two-cycle latency did not change. These optimizations were confirmed on live silicon when Jalapeño was brought up.
AMD: edits on top of an already optimized design
With AMD, OpenAI worked on the same metrics for a flagship GPU. The starting point was heavily optimized designs, and the model still found improvements of up to 14% by rewriting RTL code. According to OpenAI, that amounts to about a year of progress compressed into less than 5 months. The model rewrote circuits that compute, control operations or move data, and reasoned about them at the level of an experienced microarchitect.
The verification loop matters. Through Synopsys tools, the model confirmed that each edit did not change the block's behavior, and it assessed the tradeoff between power, performance and area. AMD has already included some of the changes in a design sent to production, and added the rest to working files for future versions. In key blocks of AMD's next flagship CPU, AMD independently confirmed a PPA gain, but it is measured in single-digit percentages.
How it was measured on RISC-V cores
To compare model generations, OpenAI took two open 32-bit cores: PicoRV32, a compact core for FPGAs, and Ibex, the embedded processor in the OpenTitan secure chip. The models got two tasks: speed up an existing core, and design a core from scratch based on written requirements. The target was throughput, measured in instructions per second.
The scoring is strict. A design got zero if it failed simulation checks or took up more than 16 times the area of the original. For the rest, OpenAI measured instructions per cycle on workloads hidden from the model and multiplied that by the maximum frequency estimated in Yosys. The best result on each core was set to 100%, and zero-score attempts counted toward the average.
With a budget of about 112 thousand tokens per run, GPT‑6.2 Astra posted the best average result on both cores, and GPT‑6.1 Sol outperformed GPT‑5.6 Sol. On PicoRV32, Astra's result is roughly double that of GPT‑6 Astra. The evaluation included 9 881 runs in total. In design from requirements, all five models scored higher than in improving an existing core: GPT‑6.2 Astra reached about 70% of the best result on PicoRV32 and roughly 75% on Ibex.
Caveats and what comes next
All figures come from OpenAI's own publication. The "up to 14%" figure refers to the best case, not an average across the whole chip. The advantage of designing from scratch is partly explained by the fact that the model may leave out performance counters and extra instructions that have little effect on the test workloads. OpenAI acknowledges that, for realism, these features should be required, area limits should be tighter, and timing should be checked after physical synthesis.
Next, OpenAI is training a model called GPT‑Synopsys together with Synopsys. It is meant to work with EDA tools (software for designing and verifying chips) the way experienced engineers do. For teams building agents, this is an example of a workable pattern: the model proposes an edit, the tool verifies it, and the metric decides.
- Nov 18, 2026One Chinese Comma Switched DeepSeek to Chinese in 94% of Responses
- Nov 18, 2026Hugging Face routes LLM requests through OpenRouter, cutting capacity errors by 46%
- Nov 18, 2026Revolut Cuts Feature Development Time by 30–60% and Hands Payment Screening to an AI Agent
- Nov 18, 2026Nvidia's data-center revenue jumps 110% to $107.6 billion: the AI pause didn't hurt
Key Questions Answered
What the models did in Jalapeño?
OpenAI began bringing GPT‑4o into Jalapeño development in early 2025. Even that model found non-trivial improvements. The vector block handles parallel computation. The savings there came from floating-point adders and multipliers, conversion functions, reduction logic and the special-function unit. The model tried alternative logic implementations and split computations across pipeline stages differently.
What are the key details on AMD: edits on top of an already optimized design?
With AMD, OpenAI worked on the same metrics for a flagship GPU. The starting point was heavily optimized designs, and the model still found improvements of up to 14% by rewriting RTL code. According to OpenAI, that amounts to about a year of progress compressed into less than 5 months. The model rewrote circuits that compute, control operations or move data, and reasoned about them at the level of an experienced microarchitect. The verification loop matters.
How it was measured on RISC-V cores?
To compare model generations, OpenAI took two open 32-bit cores: PicoRV32, a compact core for FPGAs, and Ibex, the embedded processor in the OpenTitan secure chip. The models got two tasks: speed up an existing core, and design a core from scratch based on written requirements. The target was throughput, measured in instructions per second. The scoring is strict. A design got zero if it failed simulation checks or took up more than 16 times the area of the original.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.