Google researchers published a 2021 Nature paper claiming their reinforcement-learning system could place chip macros in hours and improve power, performance, and area over human-designed layouts. The study relied mainly on five proprietary Google TPU blocks, reported limited direct comparisons, and did not fully disclose training inputs, baseline workflows, or the effect of pre-training on its claimed runtime. Critics argued that the evaluation selectively used benchmarks, metrics, and competing tools, while the fixed chip boundary meant the method did not necessarily improve total circuit area.
In 2022, Google engineer Satrajit Chatterjee drafted an alternative analysis, “Stronger Baselines,” which reportedly found established methods performed better. Google declined to publish it and terminated Chatterjee; his wrongful-dismissal lawsuit alleged retaliation after he objected to research representations he believed were misleading. A court denied Google’s motion to dismiss, finding that he had plausibly alleged unlawful retaliation, but the source does not establish that Google committed fraud. Independent UC San Diego testing in 2023 found an AlphaChip reimplementation did not outperform established or commercial placers. Nature retracted an accompanying News & Views article on September 21, 2023, added an editor’s note, and issued a September 26, 2024 addendum that supplied some missing methodological details. An ACM investigation into related criticism remained underway as of 2026.
