AI chip startup Taalas has raised $169 million to support the development of chips that have been optimized for specific AI models.

It brings the total raised by the startup to around $219m since it emerged from stealth in March 2024.

The company also unveiled its first chip, the HC1 Technology Demonstrator, which has been optimized to run the open-source Llama 3.1 8B language model. Fabricated by TSMC using its 6nm process node, Taalas claims its chips can generate more tokens per second per user than Nvidia’s H200 and B200 offerings, as well as hardware from Groq, SambaNova, and Cerebras, whilst using one tenth of the power.

By optimizing the chip for specific models, engineers are able to do away with redundant components that could otherwise slow the model down. Additionally, the company claims its approach provides a cost-effective way of developing customized chips as Taalas’ engineers only make changes to two of the layers that make up its chips.

“We basically have an architecture where we are embedding the models, and we are hard-coding the models and the weights into our what we call the mask ROM recall fabric, which is paired with an SRAM recall fabric,” explained Paresh Kharya, VP of product at Taalas, in an interview with Next Platform.

“Together, they are able to store both the model as well as do all the computations of KV cache. We have adapters and customizations – we support all of that. This design allows us to be super-dense in terms of compute and in terms of storage, and we can do compute on that storage incredibly fast, which is what drives density up and cost down.”

Taalas was founded in August 2023 by Ljubisa Bajic, a former architect at both AMD and Nvidia, and Tenstorrent co-founder, alongside engineers Drago Ignjatovic and Lejla Bajic. In October 2022, Bajic job-swapped with Tenstorrent CTO, Jim Keller, before stepping down from the company entirely in March the following year.

The company is currently developing its second chip, HC2, that will support the Llama 3.1 model with 20 billion parameters.