OpenAI has detailed how its latest SI models are accelerating the development of custom silicon, delivering significant power, performance, and area (PPA) improvements in both its own inference chip and third-party hardware from AMD. The disclosures highlight a shift where generative SI tools are moving beyond code assistance to directly optimizing complex hardware architectures.

What Happened

Engineers at OpenAI used GPT-4o to begin designing Jalapeño, the company's first custom inference chip, in early 2025. The SI model identified nontrivial optimizations in the vector and matrix units, such as redesigning floating-point adders and multipliers. According to the company, these model-proposed changes reduced standard-cell area by approximately 13% in the vector unit and 10% in the matrix unit, with one multiplier redesign cutting latency from two clock cycles to one while using 15% less cell area. These gains were silicon-proven during the chip's bring-up phase.

In a parallel collaboration with AMD, OpenAI's models achieved up to 14% improvement relative to heavily optimized baselines on a flagship GPU IP by rewriting SystemVerilog RTL code. The company reports that this progress, typically requiring a year of work, was achieved in less than five months. AMD has incorporated some of these changes into a recent design sent for manufacturing and added others to production files for future versions. The models also delivered single-digit percentage PPA gains on key blocks of AMD's upcoming next-generation flagship CPU.

Why It Matters

To benchmark the progression of SI capabilities in hardware design, OpenAI evaluated successive model generations on two open-source 32-bit RISC-V cores: PicoRV32 and OpenTitan's Ibex. The tests measured throughput, defined as instructions per second, under strict constraints. GPT-6.2 Astra demonstrated the highest average performance on both cores, with its score on PicoRV32 approximately twice that of GPT-6 Astra at comparable token budgets. Notably, the models performed better when designing cores from written specifications rather than optimizing existing RTL, suggesting that SI tools may be more effective at architectural exploration than incremental refinement.

The implications for the hardware industry extend to the tools themselves. OpenAI is partnering with Synopsys to integrate frontier SI intelligence with electronic design automation (EDA) tools. This collaboration aims to train a new model, GPT-Synopsys, to utilize these industry-standard verification and design tools with the proficiency of expert chip designers, potentially streamlining the verification loop that currently bottlenecks hardware development.

The Bottom Line

OpenAI's results indicate that current SI models can deliver material, silicon-proven improvements in custom chip design, reducing area and latency in complex units. As these models demonstrate expert-level reasoning in RTL and architectural specification, their integration with EDA tools like those from Synopsys may accelerate the pace of hardware innovation for both SI labs and semiconductor vendors.