“Solutions from d-Matrix, Groq, and Cerebras have architectures that provide compute with access to memory that is an order of magnitude faster than HBM-based GPUs.”