AMD has acquired AI chip startup Taalas, aiming to enhance inference performance and challenge Nvidia's dominance in the AI hardware market. Taalas' innovative approach involves etching AI model weights directly into silicon, a process that promises significant speedups for AI inference.

Founded in 2023, Taalas' technology differs from conventional GPUs by creating model-specific integrated circuits (MSICs). Their first test chip, the HC1, demonstrated remarkable speed, reportedly achieving 16,960 tokens per second with Meta's Llama 3.1 8B, a significant leap over existing solutions.

While this method offers substantial performance gains, it means deployed chips are fixed to a specific model, requiring a re-spin for updates. However, Taalas suggests this re-spin is less costly than starting from scratch. AMD plans to integrate Taalas' technology with its Instinct accelerators, potentially creating a disaggregated architecture for optimized AI workloads. The acquisition is expected to close in the fourth quarter, pending regulatory approval.