AMD buys Taalas, which etches models straight into silicon

AMD has acquired Taalas, a Toronto startup founded in 2023 that takes an unusual approach to inference. Instead of loading weights from HBM memory into a general-purpose GPU, Taalas etches the weights directly into the chip. The Register describes these as model-specific integrated circuits: a mask-ROM region holds the fixed weights, while a smaller SRAM region stores the KV cache and any fine-tuning adapters. Terms were not disclosed, though the report calls it a real acquisition rather than a talent grab.

The numbers are what make it interesting. Taalas said its HC1 test chip, built on a 6nm TSMC process, served Meta's Llama 3.1 8B at 16,960 tokens a second, which it claimed in February was 48 times faster than Nvidia's GPUs and 8.5 times faster than Cerebras. The next chip, HC2, aims to hold 20 billion parameters, and Taalas estimates a trillion-parameter model would need only about 50 of its accelerators against thousands of Groq LPUs. Etching a model into silicon, the company says, costs roughly 100 times less than training a frontier model in the first place.

The catch is rigidity. Once a model is burned in, you are stuck with it. Anything larger than a small adapter change requires a re-spin of the chips, although Taalas says only two metal layers need to change, which softens the cost of updates.

Why it matters

If you serve a stable model at high volume, this points to a cheaper and much faster path than renting GPUs, at the price of locking your hardware to one model version. That trade favors mature, slow-changing models, so weigh how often your production model actually changes before betting on fixed silicon.