Etched | The World's First Transformer ASIC

The article discusses the advantages of burning transformer architecture into silicon chips, a technology developed by Etched. This method allows AI models to run significantly faster and more cost-effectively than when using GPUs. The technology enables the creation of products that were previously impossible with GPUs, such as real-time voice agents that can process thousands of words in milliseconds, improved coding through tree search, and real-time generation of new content through multicast speculative decoding.

The article also highlights the potential of this technology to run future trillion parameter models. It mentions that the software stack is fully open-source and can be expanded to accommodate 100T param models. The technology also supports beam search and MCTS decoding, as well as MoE and transformer variants. The article includes 3D renders of the Sohu by Etched, presumably a product that utilizes this technology.

Key takeaways:

By burning the transformer architecture into chips, AI models can run significantly faster and cheaper than GPUs.
Products that are impossible with GPUs can be built, including real-time voice agents and better coding with tree search.
It allows for multicast speculative decoding, generating new content in real-time.
The software stack is fully open-source and expansible to 100T param models, including beam search and MCTS decoding, MoE and transformer variants.

Etched | The World's First Transformer ASIC

Key takeaways:

Comments (0)

Newsletter