Startup Taalas founded by Tenstorrent engineers claims specialized AI chip outperforms small GPU datacenters in niche workloads.

Toronto-based AI chip startup Taalas has emerged from stealth with $50 million in funding and the ambitious goal of revolutionizing the GPU-centric world dominated by Nvidia. Founded by Ljubisa Bajic, Lejla Bajic, and Drago Ignjatovic—all previously from Tenstorrent, the creator of Grayskull—Taalas is developing an automated flow for quickly turning any AI model into custom silicon.

The company is building technology to transform Transformers, SSMs, Diffusers, MoEs, and other model types into what it calls Hardcore Models, which it claims are 1000x more efficient than their software counterparts. According to Taalas, one of its chips can hold an entire large AI model without requiring external memory, and the efficiency of hard-wired computation enables a single chip to outperform a small GPU data center.

CEO Ljubisa Bajic articulated the company's vision: "Artificial intelligence is like electrical power – an essential good that will need to be made available to all. Commoditizing AI requires a 1000x improvement in computational power and efficiency, a goal that is unattainable via the current incremental approaches. The path forward is to realize that we should not be simulating intelligence on general purpose computers, but casting intelligence directly into silicon. Implementing deep learning models in silicon is the straightest path to sustainable AI."

Matt Humphrey, Partner at Quiet Capital, which led the two rounds of funding alongside advisor Pierre Lamond of Eclipse Ventures, emphasized the potential impact: "We believe the Taalas 'direct-to-silicon' foundry unlocks three fundamental breakthroughs: dramatically resetting the cost structure of AI today, viably enabling the next 10-100x growth in model size, and efficiently running powerful models locally on any consumer device. This is perhaps the most important mission in computing today for the future scalability of AI."

Taalas plans to tape out its first large language model chip in the third quarter of 2024, with the goal of making its chips available to the first customers in Q1 2025.