A company called Etched AI claims that they have created a new technology and "burned" the Transformer architecture directly into the chip.
Created the world's most powerful server dedicated to Transformer reasoning. You can run a model with trillions of parameters!
Throw NVIDIA hundreds of streets
Real-time voice proxy: It can process thousands of words in milliseconds.
Better coding and tree search: hundreds of responses can be compared in parallel.
Multicast speculative decoding: generating new content in real time.
Run the future trillion parameter model: only need one core, support all open source software stack, and can be extended to 100T parameter model.
Advanced decoding technology: including beam search and MCTS decoding.
144 GB HBM3E per chip: supports MoE and converter variants.
Etched.ai is an American chip startup company, which aims to build super-intelligent hardware and design special chips for language model reasoning.
So when it comes to entrepreneurial inspiration, he takes precautions. There is no 8-bit MAC SIMD instruction in Arm's instruction assembly. Uberti, the founder, suggested that the kernel of Arm Cortex M4 and Cortex M7 was rejected instantly, but a 16-bit instruction, which means that the operation speed of MAC SIMD is cheap during the accident. "This topic is particularly difficult, and every emperor has to face and belittle it.