Alphabet's Frozen v2 chip targets the inference overhead built into general-purpose silicon

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 neve