AMD has snapped up Taalas to give its AI inference plans another shove beyond its Helios and Cerebras tie-up.
The chipmaker signed a definitive agreement to buy Taalas only weeks after cutting a deal with Cerebras for its upcoming Helios rack-scale system.
For those not in the know, Taalas designs chips for inference workloads by baking AI models directly into silicon rather than shuffling weights from external memory.
The idea is to dodge the memory wall bottleneck. Taalas’s HC1 chip uses a proprietary digital architecture that can store four bits of data and execute maths with a single transistor.
Its design pre-computes all 16 possible products for a quantised 4-bit model weight, then uses each transistor as a physical router.
That means the circuit becomes both memory and compute, although the awkward catch is that the chip only works for the model it was built for.
X user Austin Lyons said: “WOW. AMD acquires Taalas. Has to be a high-interactivity play, but it’s clearly different than Helios + Cerebras.”
Taalas claims the HC1 setup can generate 16,000 to 17,000 tokens per second per user.
The move gives AMD another route into semi-custom AI inference, with Helios and Cerebras handling one side of the rack while Taalas could tackle decode-heavy workloads.
