Particle.news
Download on the App Store

AMD Agrees to Buy Taalas to Add Model‑Specific AI Inference Chips

The deal could slash inference costs by pairing Taalas’s chips with AMD’s Instinct GPUs, EPYC CPUs, Helios racks and ROCm software pending regulatory approval.

Overview

  • AMD announced a definitive agreement to acquire Toronto startup Taalas on Thursday, Aug. 6, and said the transaction is subject to customary closing conditions and regulatory approvals.
  • Taalas builds model‑specific integrated circuits that etch a single model’s weights into silicon; its February HC1 test chip, made on TSMC 6nm, was reported to hit roughly 16,960 tokens per second on Llama 3.1 8B in early benchmarks.
  • AMD says it will fold Taalas technology into its accelerator roadmap and system stack so Taalas chips can work with Instinct GPUs, 6th‑Gen EPYC CPUs, ROCm software and Helios rack configurations.
  • The approach trades flexibility for speed because each MSIC targets a single model and meaningful model changes require a chip respin or metal‑layer change, leaving near‑term questions about HC2 development, integration and time to market.
  • The deal follows a broader industry push toward specialized inference hardware after moves by rivals such as Nvidia and could lower per‑token operating costs for high‑volume production workloads if AMD proves performance, economics and customer adoption.