Leading EDA developers including Cadence, Synopsys, Siemens EDA, and Ansys introduced AI-enhanced software tools in the early 2020s. These initial systems utilized machine learning and reinforcement learning to optimize placement, routing, and power, performance, and area (PPA) metrics within designs defined by human engineers.
The current generation of EDA tools has evolved to include generative AI capable of writing RTL, verifying code, and suggesting design fixes. Agentic systems now automate sequences of engineering tasks, allowing AI to manage multiple tools and iterate on designs until specific targets are met without constant human intervention.
In 2023 and 2024, Cadence and Synopsys reported that hundreds of chip designs were completed using their AI-enhanced toolsets. Major tech companies have adopted these technologies with varying levels of autonomy, ranging from Google's use of AlphaChip for physical floorplanning in its TPU generations to more integrated approaches.
OpenAI utilized AI for implementation, verification, and arithmetic-circuit optimization in its Jalapeño project, achieving tapeout in nine months. The AI-assisted design improved BF16 multipliers by 56%, FP4 dot-product blocks by 21%, and FP32 accumulators by 10%, while reducing the area of matrix and SIMD units by 10% and 8% respectively.
Architect Labs developed its Redwood accelerator using an autonomous system that generated logic, RTL, and firmware based on human-defined specifications. While the system produced the design in under two weeks, the Redwood chip currently runs on an FPGA rather than being taped out as an ASIC, with performance claims based on projected designs.
