Keep pulling the thread on Anna Goldie, Azalia Miharseni.
The performance gap between AlphaChip-generated layouts and human-baseline layouts grew with each successive generation of TPU it was applied to.
Recursive Intelligence's goal is to accelerate the design of custom chips, which they believe will "bend the curve" for AI scaling laws and enable faster creation of next-generation AI.
The market for AI inference spending is over $100 billion this year and is growing rapidly.
Recursive Intelligence differentiates itself from incumbents like Cadence and Synopsys by being a "frontier AI lab" that aims to reimagine the entire chip design process, rather than replacing individual modules with AI point solutions.
Recursive Intelligence plans to release its first product within a year.
Recursive Intelligence is applying AI to the entire chip design process with the goal of transforming the industry from "fabless to designless."
The AI-driven chip placement method developed at Google produced unconventional, curved, and donut-shaped layouts to reduce wire length, power consumption, and timing violations.
A key milestone for the AlphaChip project was when a chip designed using its AI placement method was taped out, returned from fabrication, and confirmed to be working correctly.
The AI models used in the AlphaChip project demonstrated self-improvement properties, becoming more effective as they solved more chip optimization problems.
Recursive Intelligence believes synthetic data is the most promising approach for training its chip design AI, enabling a scale of data orders of magnitude beyond what any customer could provide.
Across successive generations of Google's TPU, the AlphaChip AI was adopted in an increasing area of the chip.
Google, Amazon, OpenAI in partnership with Broadcom, and Tesla are all developing custom silicon for their AI workloads.