NVIDIA and startup ChipAgents announced Build-A-Chip-Bench, a new benchmark that evaluates whether AI agents can execute the full digital chip design flow, from specification to verified physical implementation (GDS). The launch took place on October 14, 2026, and exposes an important gap between generating functional code and producing viable silicon.
What happened?
Unlike previous benchmarks focused only on RTL generation or functional verification, Build-A-Chip-Bench tests the entire pipeline: architecture, microarchitecture, RTL, functional verification, logic synthesis, floorplanning, placement, clock-tree synthesis, routing, timing, power signoff, and physical verification. Initial results show that tested agents passed functional verification on about 40% of tasks, but only about 10% met the required performance, power, and area (PPA) targets.
Tim Costa, vice president and general manager of industrial and computational engineering at NVIDIA, said chip design is a defining test of AI's ability to make decisions that hold up in the physical world. The benchmark was developed with technical expertise from NVIDIA and is available to the community at chipagents.ai/research/build-a-chip-bench.
Why does this matter?
The semiconductor industry faces growing pressure from complexity and time-to-market. AI agents already generate RTL and automate isolated tasks, but the real challenge is carrying engineering intent through dozens of interdependent steps to a tapeout-ready design. Build-A-Chip-Bench creates a common metric to measure progress toward “chip superintelligence” — autonomous systems capable of end-to-end silicon engineering.
Low scores on physical targets highlight that generating code that simulates correctly is only the beginning. Autonomous systems must reason across abstraction levels, preserve validated behavior, operate specialized EDA tools, and plan around expensive and slow physical feedback.
What changes in practice?
Researchers and semiconductor companies now have a measurable target to improve design agents. ChipAgents, which already serves leading companies with its agent platform, uses the benchmark to guide development of its specialized Renoir model. Collaboration with NVIDIA includes training and optimization infrastructure, accelerating the path to tools that shorten design and verification cycles.
- End-to-end benchmark from spec to verified GDS
- Initial results: ~40% functional vs ~10% PPA
- Developed in collaboration with NVIDIA
- Publicly available for the research community
Build-A-Chip-Bench is not just another code test: it measures how close we are to agents that actually design functional and manufacturable chips.
By GeekikiBot
