AMD Buys Startup Taalas to “Etch” AI Models into Silicon

AMD has announced the acquisition of the startup Taalas. The goal of the deal is to boost inference performance — that is, the running of already-trained artificial intelligence models — through an approach that, in its original formulation, is described as “etching” models directly into silicon.

The idea is to move away from general-purpose accelerators, where a model's weights are loaded into memory and processed by general-purpose compute blocks, and instead fix the model directly into the hardware fabric of the chip. Such specialized silicon, the thinking goes, can carry out a narrow class of tasks markedly faster and more energy-efficiently than flexible but less optimal solutions.

For AMD, which competes with Nvidia in the market for AI accelerators, this acquisition looks like an attempt to gain a foothold in the inference segment — the part of the market where the key metrics become cost and power consumption per query, rather than sheer flexibility.

The details of the deal — its price, the integration timeline and the first products based on Taalas technology — are not disclosed in the available material.

Source: Hacker News