Hardware designed around large models
MatX One uses a splittable systolic array to handle different matrix sizes efficiently. Its design combines SRAM for low latency with high-bandwidth memory for long contexts, focusing on the workloads used by frontier model developers.
Moving from design to manufacturing
The Series B supports the final stages of development and rapid manufacturing expansion. MatX had grown to a 100-person team; investors include financial funds, founders and semiconductor supply-chain companies Alchip and Marvell.