Nov 1, 2025 — by Giorgio Mendoza 2 views

After unveiling the AICore AX-M1 earlier this year, Radxa has launched the new AICore DX-M1, a compact M.2 M Key AI acceleration module designed for energy-efficient inference at the edge. The module is built around the DeepX DX-M1 processor, delivering up to 25 TOPS of INT8 performance within a 3 to 5 W power envelope.

The DX-M1 platform focuses on high-precision, low-latency inference through DeepX’s IQ8 quantization technology, which maintains FP32-level accuracy for neural networks while reducing computational overhead. The module integrates 4 GB of LPDDR5 memory and 1 Gb of QSPI NAND Flash, connected via a PCIe Gen 3 ×4 interface for high-speed data throughput.

— ADVERTISEMENT —

**AX-M1 Benchmarks **(click images to enlarge)

According to performance data from DeepX’s Model Zoo, the AICore DX-M1 achieves 3523 FPS on MobileNetV2 and 1186 FPS on ResNet50 for image classification, with 366 FPS on YOLOv8L for object detection and 223 FPS on DeepLabV3 for segmentation tasks. Energy-efficiency results reach 3242 FPS/W for MobileNetV2 and 496 FPS/W for ResNet50, demonstrating the platform’s optimization for low-power edge deployments.

AICore AX-M1 FPS per Watt(click images to enlarge)

The module measures 22 × 80 mm and adheres to the standard M.2 2280 form factor, ensuring compatibility with both industrial and embedded systems. It works with Radxa’s ROCK 5B and ROCK 5B+ single-board computers, as well as other ARM and x86 platforms featuring an M.2 M Key slot.

AICore AX-M1(click images to enlarge)

Software support includes Ubuntu and Debian Linux distributions, along with popular AI frameworks such as ONNX, PyTorch, and TensorFlow.

Further Information

The AICore DX-M1 is currently listed on AliExpress for $186.84, offered by Radxa Computer.