Updated
Updated · The Register · Aug 6
AMD Acquires Taalas, Claiming 17,000 Tokens-a-Second AI Inference
Updated
Updated · The Register · Aug 6

AMD Acquires Taalas, Claiming 17,000 Tokens-a-Second AI Inference

3 articles · Updated · The Register · Aug 6

Summary

  • Early Taalas demos showed model-specific chips generating up to 17,000 tokens a second, the performance AMD is buying to strengthen AI inference.
  • Taalas builds integrated circuits that etch individual models into silicon rather than running them on general-purpose GPUs, aiming for faster and cheaper inference.
  • AMD disclosed the acquisition as it pushes specialized AI hardware beyond its broader accelerator lineup; terms were not announced.
  • The deal adds a Toronto startup that had raised $219 million and reflects growing interest in custom inference chips as GPU limits become clearer for some AI workloads.

Insights

Can AMD's undisclosed acquisition of Taalas pay off if printing single-model chips risks massive e-waste?
Could shifting from versatile GPUs to specialized, SRAM-heavy chips finally break Nvidia's stranglehold on the AI inference market?
How will hardwired AI silicon survive if cutting-edge models become obsolete faster than a two-month manufacturing turnaround?