All news

AMD completes acquisition of Taalas, which embeds neural network weights directly into the chip

Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .

AMD gets chips where the model is "hardwired" into the die and no high-bandwidth memory is needed, but changing the model will require fabricating a new chip in about two months.

AMD has completed its acquisition of Canadian developer of specialized SI chips Taalas; the deal value could reach $2.9 billion, including $1.4 billion in contingent payments. Taalas embeds neural network weights directly into the chip's structure, combining parameter storage and computation on a single die. This removes the need for traditional high-bandwidth memory and complex packaging. In February, the startup unveiled the HC1 processor with a built-in Llama 3.1 8B model. By its own measurements, the chip generated about 17,000 tokens/s — almost 9 times faster than Cerebras at the time. Changing the model requires fabricating a new chip in about 2 months. AMD plans to use the technology alongside Instinct GPUs, as well as EPYC, ROCm and Helios.

Sources 1