MatX

MatX employs AI technologies focused on large language models, utilizing its specialized chips to optimize performance for training, reinforcement learning, and inference tasks. The architecture is tailored to support large mixture of experts (MoE) models and dense models.

Key features of MatX include high throughput exceeding 2000 output tokens per second for large models, excellent scale-up and scale-out interconnect capabilities, and a programming model that allows direct hardware control. These features cater specifically to the needs of frontier labs.

MatX is powered by advanced chip technology that combines SRAM for low latency and HBM for long-context support. This architecture allows the MatX One chip to achieve the highest throughput and lowest latency among current products.

MatX operates by utilizing high-throughput chips designed specifically for large language models (LLMs). The MatX One chip features a splittable systolic array architecture, enabling efficient processing with low latency and high throughput, particularly for training and inference tasks.