AMD deepens AI inference bet with Taalas deal as chip race heats up

AMD has acquired Toronto-based Taalas, a company that develops specialized silicon for AI inference. Taalas' approach involves integrating model weights directly into silicon, which promises to boost inference performance by an order of magnitude or more. According to The Register, this is an actual acquisition rather than an acquihire, although the terms of the deal were not disclosed.

Taalas' chips, described as model-specific integrated circuits, do not rely on high-bandwidth memory to store model weights. Instead, they etch the weights directly into the silicon. The company's first test chip, the HC1, was fabbed on TSMC's 6nm process tech and showed impressive performance in initial benchmarks. Specifically, The Register and The Register Science reported that the HC1 served Meta's Llama 3.1 8B at 16,960 tokens a second.

The acquisition is seen as part of AMD's bid to challenge Nvidia's dominance in AI hardware. As The Register noted, the deal is similar in context to Nvidia's $20 billion licensing deal with Groq last December, aiming to make high-performance "premium" inference services faster and cheaper to run.

Brief written by urgent.news from Economic Times Tech, Hacker News Best, Techmeme, The Register, The Register Science — 5 reports on this story. Machine-written — read the original for the full account.