Google just bet its inference future on a chip built for one model
2 hours ago
- Google's Frozen v2 chip hardwires parts of Gemini's architecture but keeps weights updatable to avoid obsolescence.
- The chip aims to deliver 6-10 times more tokens per watt than current Google AI chips, helping alleviate the AI compute crunch.
- This represents a shift toward model-specific silicon, similar to the evolution of Bitcoin mining from GPUs to ASICs.
- Competitors like Taalas, d-Matrix, and SambaNova are also developing specialized hardware for AI inference.
- The approach balances hyper-efficiency of hardwired architectures with flexibility for future model updates.