News
OpenAI Models Drive Advances in Chip Design for Jalapeño and AMD
OpenAI's GPT-6 Astra improved chip designs for Jalapeño and AMD, achieving up to 14% PPA gains and accelerating development timelines.
OpenAI has demonstrated significant advancements in AI-assisted chip design using its GPT-6 Astra model, optimizing both its custom inference chip, Jalapeño, and third-party hardware from AMD. These developments highlight the growing role of artificial intelligence in improving the performance, power, and area (PPA) metrics of modern silicon designs.
Improvements in Jalapeño Chip Design
OpenAI began leveraging its GPT-4o model in early 2025 to refine the design of Jalapeño, its first custom inference chip. The AI-driven approach yielded notable PPA improvements in key components:
- Warp Engine: The model optimized floating-point adders, multipliers, conversion functions, and reduction logic, reducing standard-cell area by approximately 13%. It achieved these gains by exploring alternative logic implementations and reorganizing pipeline stages to share hardware and minimize intermediate storage.
- Matrix Unit: Enhancements included redesigning multipliers and adders, which reduced cell area by about 15% each. Latency improvements were also achieved, with one multiplier’s delay dropping from two clock cycles to one. Additional optimizations in dot-product circuits and pipeline organization contributed further savings, reducing the matrix unit’s standard-cell area by 10%.
These optimizations were validated during Jalapeño’s silicon bring-up, demonstrating the practical impact of AI-assisted design.
Collaboration with AMD
OpenAI extended its AI-driven design expertise to AMD, achieving up to 14% PPA improvements on flagship GPU intellectual property. By rewriting SystemVerilog RTL code, the model compressed a year’s worth of design work into just five months. AMD has already incorporated some of these optimizations into a recent chip design sent for manufacturing, with additional improvements slated for future production designs.
The AI model demonstrated expert-level reasoning in rewriting circuits for calculations, control operations, and data movement. Synopsys tools were used to verify that each modification preserved the intended functionality while optimizing trade-offs in power, performance, and area.
Future Implications
OpenAI’s partnership with Synopsys to integrate its models into electronic design automation (EDA) tools further underscores the potential of AI in chip design. By accelerating development timelines and enhancing design efficiency, AI models like GPT-6 Astra are poised to play a transformative role in the semiconductor industry.