OpenAI says its models improved Jalapeño and AMD chip designs

OpenAI has published a post on how its models are improving chip designs. It builds on an August post about its first custom inference chip, Jalapeño, and adds results on AMD hardware, a benchmark on two open-source RISC-V cores, and a partnership with Synopsys.
Jalapeño. OpenAI began using GPT-o3 mini to help design the chip in early 2025, and says even that early model found non-trivial PPA (power, performance, area) improvements. In the vector unit, which manages and executes parallel computing tasks, savings came from floating-point adders and multipliers, conversion functions, reduction logic and the special-function unit. The models explored alternative logic implementations and ways to divide computations across pipeline stages, finding chances to share hardware and cut intermediate storage. In the matrix unit, gains accumulated across floating-point adders, multipliers and low-precision dot-product circuits. One multiplier redesign used about 15% less cell area in synthesis while cutting latency from two clock cycles to one. A separate adder redesign cut cell area by about 15% and kept its two-cycle latency. Across the broader design effort, the team estimates that model-proposed changes reduced standard-cell area by about 13% in the vector unit and 10% in the matrix unit. OpenAI says these optimizations were silicon proven during Jalapeño's silicon bring-up.
AMD. OpenAI says it partnered with AMD and saw up to 14% PPA improvement relative to heavily optimized baselines on AMD's flagship GPU IP. The model found further improvements by rewriting the SystemVerilog RTL that describes the chip's circuits. OpenAI calls this roughly a year of progress achieved in less than 5 months. AMD has already included some of the changes in a recent design it sent for manufacturing, and added others to its production design files for future versions. AMD also verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU, using designs written by the model. OpenAI says it expects further gains on GPU and CPU relative to the most optimized baseline. It also says it observed the model reason about RTL at the level of an expert microarchitect, and that the model used Synopsys tools to check that each rewrite preserved the block's behavior and to weigh power, performance and area trade-offs.
RISC-V benchmark. To track progress from one model generation to the next, OpenAI tested models on two open-source 32-bit RISC-V cores: PicoRV32, a compact core designed for FPGAs, and Ibex, the embedded processor in the OpenTitan security chip. In the first task, models got the original hardware description and were asked to raise throughput (instructions per second). Designs were checked in simulation and scored zero if they failed a check or used more than 16 times the original core's area. For passing designs, OpenAI measured instructions per clock cycle on workloads hidden from the models, took the geometric mean, and multiplied by the maximum clock frequency estimated with Yosys. The best throughput observed for each core under the same constraints scored 1.0, and the charts average scores across runs, zeros included. At a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores, and GPT-6.1 Sol outscored GPT-5.6 Sol on both.
In the second task, models received written requirements and templates showing how the core connects to the rest of the system, and chose how data moves through the core and how work is split into stages. Correctness checks, memory and cache settings and throughput objectives stayed the same. At about 112,000 tokens per run, all five models scored higher when designing a core from requirements than when improving the existing design, on both PicoRV32 and Ibex. GPT-6.2 Astra gained more on Ibex than on PicoRV32. The full evaluation covered 9,881 scored runs across both tasks, both cores and all token budgets. OpenAI notes that starting from requirements lets models omit features such as performance counters and additional instructions, which may do little for throughput on the test workloads and could explain part of the advantage. Requiring those features, setting a tighter area limit and checking timing after physical synthesis would make the test more representative of production chip design.
Synopsys. OpenAI is partnering with Synopsys to train GPT-Synopsys, a new model that uses Synopsys electronic design automation (EDA) tools the way expert chip designers do. The stated aim is to ground model development in trusted evaluators and broaden the design challenges addressed.
Key facts
- OpenAI estimates model-proposed changes cut standard-cell area on Jalapeño by about 13% in the vector unit and 10% in the matrix unit; it says these were silicon proven during bring-up.
- On AMD's flagship GPU IP, OpenAI reports up to 14% PPA improvement over heavily optimized baselines, and AMD verified single-digit percentage gains across key blocks of its next-generation flagship CPU.
- AMD has included some of the changes in a recent design sent for manufacturing and added others to its production design files.
- In a RISC-V test on PicoRV32 and Ibex (about 112,000 tokens per run), GPT-6.2 Astra scored highest on both cores, and all five models did better designing from written requirements than improving existing designs.
- OpenAI is partnering with Synopsys to train GPT-Synopsys, a model meant to use Synopsys EDA tools the way expert chip designers do.
Why it matters
Chip design is the base layer of the AI stack, and OpenAI reports that its models now contribute to production designs rather than only to toy benchmarks. The claims reach beyond its own silicon: AMD has included some changes in a design sent for manufacturing. OpenAI describes the GPU progress as roughly a year of progress in less than 5 months. The Synopsys tie-up points to the next step, a model trained to operate the EDA tools engineers already use.
Who it affects
Chip design teams and the EDA tools market are the most direct audience, since the work runs through Synopsys tools and a model trained on them. AMD is a named partner on both GPU and CPU work. OpenAI's own Jalapeño team is the first user. The RISC-V results also bear on anyone tracking how each new model generation performs on hardware design tasks.
How to use it
The post describes internal work and partnerships rather than a tool to adopt. GPT-Synopsys is still being trained, and the source does not say when it will be released. For now the practical takeaway is the method: models rewrite RTL, and Synopsys tools check that behavior is preserved and measure power, performance and area.
How solid is it
This is a first-party OpenAI post. Some claims have outside backing: AMD verified the CPU gains and has used some of the GPU changes in a design sent for manufacturing, and OpenAI says the Jalapeño optimizations were silicon proven. Still, no independent verification beyond AMD's own is mentioned, and the post gives percentages only, with no absolute watts, MHz or mm2 figures. The Jalapeño area savings are the team's own estimates. The 14% GPU figure is an upper bound ('up to'), and the source does not say which blocks it applies to. The CPU gains are only 'single-digit percentage'. No numeric RISC-V scores appear in the text, only rankings.
Risks and caveats
OpenAI itself says the RISC-V test is not fully representative of production design. Models designing from requirements could omit features like performance counters and extra instructions that matter little on the test workloads, which may explain part of their higher scores. A tighter area limit, required features and timing checks after physical synthesis would make the test closer to real chip design. The AMD GPU and CPU products are described only as 'flagship GPU IP' and an 'upcoming next-generation flagship CPU'. The further gains OpenAI expects are a forward-looking expectation, not a result.
“We’ve observed the model reason about RTL at the level of an expert microarchitect.”
— OpenAI