Google develops Frozen v2 Gemini AI chip

Hardwiring model architectures is a bold bet on hardware-software co-design, offering massive efficiency gains at the cost of long-term flexibility. Direct hardware implementation of model layers eliminates memory access overheads, which are the primary source of energy consumption in inference. While an order-of-magnitude efficiency improvement is crucial for scaling Gemini services sustainably, designing a chip for deployment in 2028 risks committing to design architectures that could become obsolete before launch.