Radxa AICore DX-M1M Edge AI Acceleration Module Hits 25 TOPS

Radxa AICore DX-M1M edge AI acceleration module

Built around the DeepX DX-M1M NPU, the Radxa AICore DX-M1M is a petite, low-power M.2 edge AI acceleration module that employs merely 3W of power to accomplish up to 25 TOPS (INT8) of AI performance. One of the foremost obstacles for embedded engineers in edge computing is to gain high-performance neural network inference within confined power and space. This need is met by the Radxa AICore DX-M1M, a compact M.2 2242 AI acceleration module co-developed with DEEPX that is intended to offload compute-intensive workloads from embedded systems and single-board computers (SBCs). The DeepX DX-M1M SoC, designed specifically for edge AI inference, is the central component of the AICore DX-M1M. The DX-M1M offers better space efficiency and optimized power consumption without compromising performance when compared to the previous AICore DX-M1 (M.2 2280 form factor).

With support for both x86 and ARM-based systems, the AICore DX-M1M has been integrated over extensive platform compatibility. Raspberry Pi 5, Radxa ROCK 5A, ROCK 5B, ROCK 5B+, and ROCK 5 ITX are among the platforms that are compatible. In a typical M.2 M + B Key (PCIe Gen3 ×2) form factor, this AI processor combines QSPI flash and onboard memory. The module offers high-performance AI and ML capabilities without going over the power budget and is intended for use with industrial robot arms, autonomous mobile robots (AMR), edge servers, drones, and AIoT devices.

Radxa AICore DX-M1M Front and Back View

AICore DX-M1M Specifications:

• AI Engine: Powered by the DeepX DX-M1M NPU, delivering up to 25 TOPS of AI compute performance
• Memory: Integrated 1GB LPDDR4X running at 4266 MT/s (scalable up to 8GB as per DeepX support)
• Storage: Equipped with 1Gbit QSPI NAND/NOR flash
• Module Interface: PCIe Gen3 x2
• Host Interface: PCIe Gen 3.0 x4 host interface (backward compatible with Gen 1/2 and configurable to x1/x2 lanes) via M.2 M + B Key
• Power Consumption: Typical operating power of just 3W
• Mechanical Details:
  • Form Factor: M.2 2242 (42 x 22 mm)
  • Compatibility: Can be used in M.2 2280 slots with an adapter
• Operating Temperature
  • -25°C to 65°C: Maintains stable, non-throttled performance
  • 65°C to 85°C: Thermal protection enabled with performance throttling

Radxa AICore DX-M1M installed on a Raspberry Pi 5

In terms of the software support, the DEEPX DXNN SDK, which enables model compilation, optimization, and hardware-accelerated inference on the DEEPX NPU, is used by the Radxa AICore DX-M1M. Through the DX-COM compiler, it converts models from PyTorch, ONNX, TensorFlow, and Keras to the proprietary DXNN format. The DX-All Suite, an extensive toolchain that streamlines model conversion and runtime deployment, supports the AICore DX-M1M. The getting started guide possesses a fuller explanation. 

DX-M1M Installation

The Radxa AICore DX-M1M, which is listed at $85.00 from Arace Tech and $97.34 on AliExpress, offers an attractive price-to-performance ratio for edge deployments that prioritize compute density and energy efficiency. For any additional information, perceive the product page. 

Images used courtesy of radxa and arace.tech