AMD acquires Taalas, a startup that bakes AI models directly into silicon

AMD acquires Taalas to embed AI model weights directly into custom silicon chips. The startup's hardware achieves speeds exceeding 16,000 tokens per second for Llama 3.1-8B by hard-coding models into silicon, though it sacrifices flexibility. This acquisition signals a major shift toward application-specific hardware as chipmakers race to drastically accelerate LLM inference performance.

Opening Kapyn…