AMD buys Taalas to hardwire AI model weights directly into silicon
AMD acquired Toronto-based startup Taalas to boost AI inference speed by baking model weights straight into custom silicon chips.
AMD has acquired Taalas, a Toronto-based AI chip startup founded in 2023, to integrate its unconventional model-specific hardware into its compute ecosystem. Financial terms of the acquisition were not disclosed by AMD.
Unlike traditional GPUs or dataflow architectures like Groq’s LPU and Cerebras wafer-scale accelerators, Taalas builds Model-Specific Integrated Circuits (MSICs). The startup hardwires language model weights directly into the silicon architecture using an inversion approach, splitting its processor layout between mask-ROM and SRAM. In February, the startup showcased its 6nm TSMC-built HC1 test chip, which processed Meta Llama 3.1 8B at 16,960 tokens per second—a figure reported at the time to outpace standard Nvidia GPUs by 48 times and Cerebras accelerators by 8 times.
AMD plans to pair its Instinct-based Helios processors with Taalas-derived chips in a disaggregated architecture. In this setup, resource-heavy request processing handles on graphics chips, while token generation offloads to the Taalas-based accelerators. The startup is currently planning a second-generation HC2 chip targeting 20 billion parameters, which it estimates could run on just 50 accelerators.
Source: AMD Press Release
Source: habr
Comments
Help shape the next version: Add context or suggest a correction. AI review can add points toward a rewrite. Reviews and updates may take time; a full meter does not guarantee a new version.