AI challenger Cerebras assembles modular supercomputer ‘Andromeda’ to speed up large language models

The building-block approach is able to dramatically reduce compute time compared to GPUs in direct proportion to how many machines are joined together.