AI news ·
OpenAI models improve chip designs for custom and third-party hardware
OpenAI's models cut standard-cell area by up to 13% on its custom Jalapeño inference chip and helped AMD achieve up to 14% improvement on flagship GPU IP.

OpenAI's models are now contributing expert-level design improvements to production chips and open-source processor cores, with the company reporting silicon-proven gains on its custom Jalapeño chip and a new collaboration with AMD that delivered up to 14% improvement on flagship GPU IP. The work demonstrates that frontier models can reason about hardware description code at the level of a skilled microarchitect, compressing roughly a year of design progress into less than five months.
Silicon-proven gains on Jalapeño
OpenAI engineers began using GPT-o3 mini to assist with Jalapeño, the company's first custom inference chip, in early 2025. The model proposed changes that reduced standard-cell area by about 13% in the vector unit and 10% in the matrix unit. These optimizations were confirmed during silicon bring-up.
In the vector unit, 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 opportunities to share hardware and reduce intermediate storage. In the matrix unit, one multiplier redesign cut cell area by roughly 15% while reducing latency from two clock cycles to one. A separate adder redesign reduced cell area by about 15% while keeping its two-cycle latency. Further improvements to dot-product logic and pipeline organization added savings across the block.
AMD partnership shows rapid progress
OpenAI has partnered with AMD on power, performance, and area (PPA) improvements. On AMD's flagship GPU IP, the model identified gains of up to 14% relative to heavily optimized baselines by rewriting SystemVerilog RTL. AMD has already included some of these changes in a design sent for manufacturing and added others to its production design files for future versions.
"This represents roughly a year of progress achieved in less than 5 months," OpenAI said. The model used Synopsys tools to verify that each rewrite preserved block behavior and to evaluate trade-offs in power, performance, and area. AMD also verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU using model-written designs.
RISC-V benchmarks show generational improvement
To track capability progression, OpenAI tested successive model generations on two open-source 32-bit RISC-V cores: PicoRV32 and OpenTitan's Ibex. The models were asked to increase throughput - the number of instructions executed per second - with designs that failed correctness checks or exceeded 16 times the original area scoring zero.
At a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra achieved the highest average score on both cores. On PicoRV32, its score was roughly double GPT-6 Astra's. When models designed a core from written requirements rather than optimizing existing RTL, all five tested models scored higher. GPT-6.2 Astra averaged about 0.75 near 112,000 tokens on Ibex when starting from a specification. The full evaluation included 9,881 scored runs across both tasks, both cores, and all token budgets.
Synopsys partnership and next steps
OpenAI is working with Synopsys to train a new model, GPT-Synopsys, to use electronic design automation tools the way expert chip designers do. The goal is to ground model development in trusted evaluators and broaden the design challenges the models can address. For teams working at the intersection of hardware and AI, understanding how models interact with RTL and verification flows is becoming directly relevant to engineering workflows. Professionals exploring AI code generation courses or AI design courses may find the chip design domain a useful case study in how frontier models handle constrained optimization problems with real-world verification requirements.
Why this matters for IT and development professionals
The results signal that AI-assisted hardware design is moving from research to production. For developers and IT architects, the immediate takeaway is that models can now reason about low-level hardware description languages and propose optimizations that survive physical synthesis and silicon validation. This does not replace chip designers, but it compresses iteration cycles. Teams working on embedded systems, FPGA deployments, or custom silicon should watch how EDA tool integration evolves - the same verification workflows you already use may soon include model-generated candidates worth testing.