NewsNTech
General-purpose AI accelerators are designed to execute any model architecture, and that breadth carries a cost: silicon area and power both go toward flexibility that a dedicated model never uses.
Alphabet is reported to be developing a chip called "Frozen v2" that would embed parts of Gemini's architecture directly into the hardware, and Alphabet's stock rose on the news.
The constraint that model-specific silicon addresses Every AI chip must choose how much of its compute to make programmable.
A fully programmable chip can run any model but spends cycles fetching, decoding, and routing instructions for operations that a specific model will execute in a fixed pattern every time.
Keep reading