MemryX

MemryX, established in 2019, focuses on edge AI processing with its flagship product, the MX3 AI accelerator chip. The MX3 chip delivers over 5 TFLOPs of performance and supports activations in Bfloat16 and weights in 4, 8, and 16 bits. It features a batch size of 1, stores approximately 10 million parameters on-die, and has an average power consumption of about 1.0W. The chip supports USB 2.0/3.x and PCIe Gen 3 I/O, making it easily integrable into existing systems. Designed for edge AI applications, the MX3 chip is suitable for various markets including Industrial 4.0 & Robotics, Automotive, IoT, Metaverse, Smart Vision, and Edge Computing. It was notably demonstrated integrated into the Lenovo ThinkCentre neo Ultra desktop PC at CES 2024, marking it as the first third-party AI accelerator card used in a PC. MemryX’s architecture employs a proprietary, configurable native dataflow architecture with at-memory computing, which eliminates data movement bottlenecks and supports future hardware and AI models. The company’s MX-SDK allows developers to create systems and applications that optimally utilize MemryX AI accelerator chips, supporting 1-click compilation and mapping of trained AI models. MemryX has offices in Ann Arbor, Michigan; Taipei, Taiwan; Hsinchu, Taiwan; and Bangalore, India. The MX3 chip can be used in applications such as logistics and warehousing robots, manufacturing inspection, in-cabin automotive systems, smart retail, agriculture, drones, AR/VR, surveillance, smart cities, and video management systems. It supports pipelined execution for high performance and low latency and can be scaled by cascading multiple chips for increased performance or lower latency.