AMD has agreed to buy Taalas, a Toronto startup whose inference chips are architected around specific AI models, adding a specialized accelerator line to its GPU-heavy AI roadmap.
Taalas was founded in 2023 by Ljubisa Bajic, a former Tenstorrent CEO and ex-AMD executive, and raised $169M in February. The company optimizes inference dataflows to remove the compute and memory bottlenecks of general-purpose architectures, AMD said.
EE Times reports the Taalas chips will likely run alongside AMD’s Instinct GPUs as LLM decode accelerators, cutting the latency and power cost of generating tokens. Bajic and his team will join AMD’s AI group under Vamsi Boppana.
AMD says the technology will slot into its full-stack platform spanning Helios rack-scale systems, EPYC server CPUs and ROCm software. Terms were not disclosed.
The pickup gives AMD dedicated inference silicon at a moment when rivals are pushing specialized parts for low-latency AI workloads that general-purpose GPUs handle inefficiently.
