AMD Acquires AI Chip Startup Taalas to Etch Models Directly into Silicon
AMD has acquired Taalas, a Toronto-based AI chip startup that specializes in etching machine learning models directly into silicon. This acquisition aims to advance AMD's AI inference capabilities by eliminating the need for external memory like High Bandwidth Memory (HBM) to store model weights. By hardwiring model weights into the silicon, this technology bypasses the traditional memory bandwidth bottleneck that limits AI inference performance. This acquisition could allow AMD to offer highly efficient, specialized chips for stable, high-volume AI models, challenging Nvidia's dominance in the inference market. Taalas's approach creates "hardcore models" that do not rely on external DRAM or software, significantly reducing power consumption and latency. However, a key trade-off is the lack of flexibility, as the etched models cannot be updated or modified once the silicon is manufactured.
## BACKGROUND
Traditional AI inference relies on general-purpose processors like GPUs, which must constantly fetch model weights from external memory (such as HBM or DDR) during computation. This process is highly energy-intensive and limited by memory bandwidth, a challenge known as the "memory wall" in AI hardware design.