“MatX's chip architecture combines HBM for inference data and SRAM for model weights, aiming to deliver both high throughput and low latency.”