MemryX MX3 — M.2 AI Accelerator Module

Specifications

Overview

The MemryX MX3 is an M.2 2280 M-key AI accelerator module that adds a dedicated AI 'artificial brain' to any PCIe Gen 3 host. Each module carries up to four MemryX accelerator chips, delivering 6 TFLOPS at 1 GHz per chip — up to 96 TFLOPS pipelined — while consuming just 0.6–2W per chip.

Processor-agnostic across x86, ARM and RISC-V, the MX3 supports 4-bit and 8-bit weights and standard frameworks (ONNX, PyTorch, TensorFlow, Keras, TF Lite) on Ubuntu, Windows and Android. Its -40°C to +85°C operating range and extreme power efficiency make it a compelling option for power- and thermal-constrained edge AI deployments.

With capacity for up to 80 million 4-bit parameters, the MX3 targets real-time inference on data streams — smart cameras, industrial inspection, and embedded systems where a full GPU would be overkill in cost, power and thermal budget.