AMD buys Taalas to speed AI inference with model-specific chips

AMD has acquired Toronto-based AI chip startup Taalas, a move aimed at improving inference performance as it competes with Nvidia in AI hardware. Terms were not disclosed, and the deal is expected to close in the fourth quarter, subject to regulatory approval.

Taalas builds model-specific integrated circuits that etch model weights directly into silicon instead of storing them in HBM. Founded in 2023, the startup demonstrated its HC1 test chip on TSMC’s 6nm process, serving Meta’s Llama 3.1 8B at 16,960 tokens a second, which was described at launch as 48x faster than Nvidia’s GPUs and 8.5x faster than Cerebras’ accelerators.

The company’s next HC2 chip is due out this summer and targets 20 billion parameters. AMD is expected to pair Taalas-based accelerators with Instinct-based Helios racks, potentially using GPUs for prompt processing while offloading token generation to the new silicon.

The approach carries a major tradeoff: once deployed, chips are tied to a specific model, and larger changes require a re-spin. Taalas has said only two layers of metal need to change for a new model, making the process less costly than starting over.