Here's what published — published is not granted. Application US20240370693A1, "Full-Stack Hardware Accelerator Search," published November 7, 2024, lists inventors Dan Zhang, Azalia Mirhoseini, Anna Goldie, and Ebrahim Songhori — names associated with Google's machine-learning-for-chip-design research. The CPC codes are G06N 3/04 (architecture) and G06N 5/04 (inference engines).
The mechanism is search over designs. Designing an AI accelerator means choosing thousands of interacting parameters — compute layout, memory hierarchy, dataflow — and the combinations explode. "Full-stack" search means automatically exploring that joint design space, across both the hardware and the software/compiler stack that runs on it, to find configurations that perform well. In effect, it's using machine learning and automated search to design the machines that run machine learning.
That recursive quality is the strategic story: the same labs building large models are building tools to automate the design of the hardware those models need. Owning a full-stack accelerator-search method protects a meta-level capability — not a chip, but the method for discovering good chips. The inventor list ties it to a well-known line of learning-to-design research.
Because this is a publication, treat the claims as sought. The allowed claims, if a grant issues, will define the scope, and there is nothing enforceable yet. The document signals where the design-automation research is heading, not a settled right.
The takeaway: US20240370693A1 is a published marker at the most recursive edge of the field — AI methods that search the design space of AI hardware — and a reminder that the published-versus-granted line is exactly where an exciting title meets the discipline of actual claims.