AMD Acquires Taalas to Pivot Toward Model-Specific Silicon

AMD has acquired Toronto-based AI chip startup Taalas in a move to challenge Nvidia's hardware dominance, as The Register first reported. Unlike conventional GPUs or dataflow architectures, Taalas utilizes "model-specific integrated circuits" (MSICs) that etch model weights directly into the silicon rather than relying on high-bandwidth memory (HBM).

Early performance benchmarks for Taalas' first test chip, the HC1, showed it serving Meta’s Llama 3.1 8B at 16,960 tokens per second—a rate The Register reports was 48x faster than Nvidia GPUs and 8.5x faster than Cerebras accelerators. The hardware consists of a mask-ROM recall fabric for etched weights and an SRAM recall fabric for fine-tuning adapters and KV caches.

AMD plans to integrate this technology into its Instinct-based Helios racks. Vamsi Boppana, AMD’s SVP of AI, stated the company is creating a full-stack platform to provide "the right compute solutions for every AI workload." The proposed architecture would likely offload token generation to Taalas accelerators while GPUs handle compute-heavy prompt processing.

While the approach offers significant power and space efficiency—potentially requiring only 50 accelerators to support a trillion-parameter model—it introduces a hardware rigidity. Because weights are etched, any major model update requires a chip re-spin. Taalas claims this process is mitigated because only two metal layers need to be changed.

In an interview with The Next Platform, Taalas suggested that etching weights into silicon costs 100x less than training a frontier model. AMD may target major Instinct customers, such as Meta, Anthropic, and OpenAI, for these deployments.