
Fixed chip beats general GPU by 690x on language model tokens
The authors co-design a fixed-architecture processor and a language model, holding weights and KV cache on-die, to approach the arithmetic energy floor of token generation. At 5 nm, the design decodes Qwen3-1.7B tokens at 0.52 mJ, 1.8x the floor and 690x less energy than a B300, while serving twelve users 110x faster.















































