MatX

MatX is an AI chip startup focused on designing processors optimized for training large language models (LLMs). The company aims to deliver performance 10x superior to NVIDIA GPUs, emphasizing cost-efficiency over latency in its chip design. For models with 70 billion parameters, MatX promises sub-100 millisecond latency per token.

MatX provides developers with low-level hardware control, enabling performance-enhancing computation customizations. Its streamlined architecture, free of unnecessary GPU components, enables integration of additional AI-specific circuits. The company’s interconnect technology further enhances scalability, improving performance across large AI clusters.

MatX claims its chips can support LLMs with up to 10 trillion parameters, while smaller models (e.g., 70 billion parameters) can be trained in days or weeks. MatX expected to launch its first product in 2025 (December 2024).