designed to provide efficient AI processing capabilities for compact computing systems. The module targets edge computing applications, where power efficiency and compact design are critical.
Packing four MemryX MX3 AI accelerator chips onto a standard M.2 2280 form factor, enables the module to be easily integrated into systems equipped with a PCIe Gen 3 M.2 slot. Each MX3 chip is said to deliver 6 TOPS (Tera Operations Per Second), offering a total of up to 24 TOPS of compute power while consuming just 6 to 8 watts of power. It also supports a range of data formats, including 4-bit, 8-bit, 16-bit weights, and BFloat16. Notably, the module operates without active cooling, relying on an included passive heatsink to manage thermal performance.
Phoronix has tested the MemryX MX3 M.2 module, evaluating its performance and usability as an AI accelerator. Installed in a system with an M.2 PCIe Gen 3 slot running Ubuntu 24.04 LTS, the module proved straightforward to integrate thanks to MemryX’s open-source drivers and developer tools. The module’s 24 TOPS performance from its four MemryX MX3 chips is said to be sufficient for various inference workloads, particularly those optimized for 8-bit weights.
The review also highlighted the module’s strong software compatibility, supporting frameworks like TensorFlow and ONNX, and its efficiency in running small to medium-sized AI models. Each MX3 chip supports up to 10.5 million 8-bit parameters, with the module’s four chips combined handling a total of up to 42 million parameters. This limitation arises from the absence of onboard DRAM. MemryX plans to address that with a new PCIe card in 2025 that will feature more MX3 AI chips onboard.
The MX3 M.2 module is available for $149 USD, making it an affordable option for developers and organizations seeking to add AI processing capabilities to edge devices. MemryX has also announced that the module will be showcased at the upcoming Consumer Electronics Show (CES) 2025 in Las Vegas where it plans to demonstrate the MX3’s performance across various real-world applications, further highlighting its versatility.
