Toronto-based AI chip startup Taalas has emerged from stealth with $50 million in funding and a lofty goal to revolutionize the GPU-centric world dominated by Nvidia.
Founded by Ljubisa Bajic, Lejla Bajic, and Drago Ignjatovic, all from Tenstorrent (creators of Grayskull), Taalas is an automated flow for rapidly converting AI models such as transformers, SSMs, diffusers, MoEs, etc. into custom silicon. We are developing. . The company claims that the resulting hardcore model is 1000 times more efficient than its software equivalent.
The company also claims that one of its chips can hold an entire large AI model without the need for external memory, and the efficiency of hard-wired computation allows a single chip to outperform small GPU data centers. It also states that performance can be achieved.
Casting intelligence directly into silicon
“Artificial intelligence is like electricity, an essential good that must be made available to everyone. Commoditizing AI will require a 1000x increase in computing power and efficiency; A goal that cannot be achieved with a step-by-step approach. The way forward is to: “Instead of simulating intelligence on general-purpose computers, we should cast intelligence directly into silicon. Implementing deep learning models in AI is the most direct path to sustainable AI,” said Ljubisa Basic, CEO of Taalas.
“We believe Taalas’ direct-to-silicon foundry will enable three fundamental breakthroughs: dramatically resetting the cost structure of today’s AI; This is likely the key to future scalability of AI: to realistically enable 10-100x growth in AI, and to run powerful models locally and efficiently on any consumer device. This is the most important mission in computing today, and we are proud to support the efforts of this incredible n-of-1 team,” said Matt Humphrey, partner at Quiet Capital, who led both rounds. says Mr. An advisor to Eclipse Ventures, he is co-funding with Pierre Lamond.
Taalas says it aims to develop its first large-scale language model chip in the third quarter of 2024 and deliver the chip to its first customers in the first quarter of 2025.