June 25th, 2024

Sohu: The First Transformer ASIC

Etched secures $120 million to develop Sohu, a transformer ASIC enhancing AI model performance. Sohu enables real-time voice agents, rapid text processing, and trillion-parameter models, revolutionizing AI processing with advanced features.

Read original articleLink Icon
Sohu: The First Transformer ASIC

Etched has secured $120 million in funding to develop Sohu, the world's first transformer ASIC. By embedding the transformer architecture into their chips, Etched aims to enhance AI model performance significantly, making them faster and more cost-effective than GPUs. Sohu promises to enable tasks like real-time voice agents, rapid text processing, and advanced content generation. The ASIC is designed to support tomorrow's trillion-parameter models, offering open-source software and scalability up to 100T parameter models. With features like beam search, MCTS decoding, and support for MoE and transformer variants, Sohu is positioned to revolutionize AI processing. Etched's innovative approach marks a significant step towards unlocking new possibilities in AI hardware acceleration.

Link Icon 6 comments
By @mysterEFrank - 5 months
Do they bake in the actual weights or the architecture? If it's just the architecture I don't understand where a speedup that considerable can come from.
By @galaxyLogic - 5 months
Does using ASICs mean something like using FPGAs? If so AMD should be well-prepared for this architecture ever since their Xilinx acquisition.

But isn't it the case that for LLMs what is really needed is lots of fast memory? ASICs can't help much there can they?

By @Zaheer - 5 months
The results sort of speak for themselves - custom ASIC's are the way of the future. How hard is it though for Nvidia to design a custom ASIC like this?
By @ChrisArchitect - 5 months