‘A single chip that outperforms a small GPU data center’: Yet another AI chip company wants to challenge Nvidia’s GPU-centric world – Taalas wants to have super-specialized AI chips
Toronto-based AI chip startup Taalas has emerged from stealth with $50 million in funding and the lofty goal of revolutionizing the GPU-centric world dominated by Nvidia.
Founded by Ljubisa Bajic, Lejla Bajic and Drago Ignjatovic, all formerly of Tenstorrent (the creator of Grayskull), Taalas develops an automated flow to quickly convert any AI model – Transformers, SSMs, Diffusers, MoEs, etc. – into modified silicon . The company claims that the resulting hardcore models are 1000x more efficient than their software counterparts.
The startup also says that one of its chips can house a very large AI model without the need for external memory, and that the efficiency of wired computation allows a single chip to outperform a small GPU data center.
Pouring intelligence directly into silicon
“Artificial intelligence is like electrical energy – an essential commodity that must be made available to everyone. Commoditizing AI requires a thousandfold improvement in computing power and efficiency, a goal unattainable with current incremental approaches. The way forward is to realize “We should not simulate intelligence on general-purpose computers, but infuse intelligence directly into silicon. Implementing deep learning models in silicon is the most direct path to sustainable AI,” says Ljubisa Bajic, CEO from Taalas.
“We believe the Taalas ‘direct-to-silicon’ foundry unlocks three fundamental breakthroughs: dramatically resetting the current cost structure of AI, making the next 10-100x growth in model size feasible, and efficiently running high-performance locally models on any consumer device. This is perhaps the most important mission in computing today for the future scalability of AI. And we are proud to support this remarkable n-of-1 team in doing so,” said Matt Humphrey, partner at Quiet Capital, who led the two rounds of funding together with Pierre Lamond, advisor at Eclipse Ventures.
Language says it will launch its first major language model chip in the third quarter of 2024 and aims to make its chips available to the first customers in the first quarter of 2025.