AUTO-UPDATED

OpenAI pledges to add Astra security as Anthropic loosens Fable's leash

AMD has acquired AI chip startup Taalas to accelerate inference performance by etching models directly into silicon, achieving speeds of up to 17,000 tokens per second in demonstrations.

Key Points

  • AMD purchased Taalas to integrate model-specific circuits into hardware for enhanced AI inference efficiency.
  • Early technical demonstrations show the new silicon architecture processing up to 17,000 tokens per second.
  • The acquisition aims to optimize performance by moving AI model execution from software to dedicated hardware layers.
  • This move reflects a broader industry trend of hardware manufacturers seeking to improve AI processing speeds through custom chip design.

Why it Matters

This acquisition signals a strategic shift toward hardware-level optimization to overcome the latency and energy constraints of current AI inference models. By embedding models directly into silicon, AMD could significantly lower operational costs and improve performance for enterprise-scale AI deployments.
Theregister.com Published by Thomas Claburn
Read original