AMD acquires AI chip startup Taalas

// 1h agoFUNDING MNA

AMD acquires AI chip startup Taalas

AMD has acquired a Toronto-based startup that bakes Large Language Model (LLM) weights directly onto silicon instead of relying on high-bandwidth memory (HBM). Early test chips demonstrated running Llama 3.1 8B at around 17,000 tokens per second—roughly 48 times faster than standard NVIDIA GPUs. However, because the model logic is physically hardwired, updating the model requires a complete chip re-spin.

// ANALYSIS

Etching weights onto silicon trades traditional software adaptability for staggering inference performance gains, marking a bold shift toward hyper-specialized AI hardware.

–Delivering 17k tokens/sec unlocks entirely new real-time voice and agentic application paradigms.
–The severe drawback of hardware lock-in makes this approach viable primarily for frozen, high-volume foundational models.
–AMD's acquisition underlines a growing push to bypass traditional memory-bandwidth bottlenecks through unconventional chip design.

// TAGS

taalasamdacquisitionllmsiliconhardwareinferencellama-3.1semiconductors

DISCOVERED

1h ago

2026-08-08

PUBLISHED

2h ago

2026-08-08

RELEVANCE

9/ 10

AUTHOR

bsormagec