AMD Agrees to Buy Taalas, Which Builds Chips Hardwired for a Single AI Model

AMD has reached a definitive agreement to acquire Taalas, a Toronto startup that makes specialized AI inference chips, the company announced on Aug. 6. AMD didn’t disclose terms. The deal is subject to customary closing conditions and regulatory approvals.

Founded in 2023, Taalas designs hardware around a specific model. Instead of loading weights from memory such as HBM and running them on programmable compute, it burns the model into the chip, ServeTheHome explained. Changing a model means changing the chips.

The HC1 demonstrator

Taalas’ current hardware is the HC1, a technology demonstrator shown running Meta’s Llama 3.1 8B. Taalas claims up to 17,000 tokens per second per user and lists the chip as TSMC 6nm, with an 815-square-millimeter die and 53 billion transistors. It compares the HC1 with Nvidia’s H200 and B200 and with Groq, SambaNova and Cerebras systems, but those figures are Taalas’ own measurements.

Larger models would need several reticle-size chips to hold the whole model. Taalas says changing weights, matrix dimensions and other key parameters only requires changing two mask layers, which keeps the number of expensive masks down even when a model spans many chips.

How AMD plans to use it

AMD said Taalas’ technology will complement its Helios rack-scale systems, Instinct GPUs, EPYC CPUs and ROCm software, and that it plans to integrate the technology into its accelerator roadmap and build system-level products with Instinct GPUs.

“AMD is building a full-stack AI platform that gives customers the flexibility to deploy the right compute solutions for every AI workload,” said Vamsi Boppana, senior vice president of AMD’s Artificial Intelligence Group. “Taalas’ technology and world-class engineering team strengthen our AI portfolio by delivering differentiated inference performance and efficiency.”

“We founded Taalas to rethink AI inference from the ground up by building the hardware around the model,” said Ljubisa Bajic, Taalas’ co-founder and CEO. “Joining AMD will give us the scale, engineering resources and global reach to accelerate our innovation.”

The trade-off is flexibility for efficiency. A rack built around one model can’t be retargeted when the model changes, so ServeTheHome sees the approach fitting stable, high-volume inference, the part of the market AMD says is growing fastest. The open question is how quickly the demonstrator becomes a shipping product.