AMD Acquires Taalas to Integrate AI Models Directly Into Silicon
AI-generated image · US National Wire
The acquisition of the Toronto-based startup aims to boost inference performance through model-specific integrated circuits that etch weights into hardware.
AMD has acquired Taalas, a Toronto-based AI chip startup, in an effort to challenge Nvidia's hardware dominance, as The Register first reported. According to reporting from The Register, Taalas utilizes a process that etches model weights directly into silicon, creating what are known as model-specific integrated circuits (MSICs). Unlike conventional GPUs, these chips do not rely on HBM to store model weights.
Early performance data indicates significant efficiency gains. The Register reports that Taalas' first test chip, the HC1—fabbed on TSMC’s 6nm process—served Meta’s Llama 3.1 8B at 16,960 tokens per second. This rate was 48x faster than Nvidia GPUs and 8.5x faster than Cerebras accelerators at the time of the February announcement. Taalas' upcoming second-generation HC2 chip aims to support 20 billion parameters.
AMD plans to integrate this technology into its Instinct-based Helios racks. Vamsi Boppana, AMD’s SVP of AI, stated that the company is developing a full-stack platform to provide customers with flexible compute solutions for various AI workloads. The Register notes that AMD may use a disaggregated architecture, utilizing GPUs for prompt processing while offloading token generation to Taalas accelerators.
While the technology offers massive speed increases, it lacks the flexibility of general-purpose hardware. Because weights are etched into the silicon, any significant model change requires a chip re-spin. However, Taalas claims this process is less costly as it only requires changing two layers of metal. In a February interview with The Next Platform, Taalas suggested that etching weights into silicon is 100x less expensive than training a frontier model.