The company claims its approach can speed up inference performance by 1000x over general purpose AI chips .
CEO is Tenstorrent co-founder Ljubisa Bajic. Investors include Quiet Capital, Fidelity and Pierre Lamond.
Taalas’ first chip calked HC1 (pictured) achieves over 16,000 tokens per second per user on Llama3.1-8B.
It says it can make a chip for any AI model in a couple of months. It can do this because its customisations only require two metal layers on a previously fabbed chip which TSMC can turn round on its N6 process in two months.
Taalas is working up from less complex to more complex models and expects to be able to handle leading-edge models by the end of the year.
