New Edge AI Accelerator Offers 8GB LPDDR4X RAM and 8K Ultra HD Video Support

Radxa AICore AX-M1 Edge AI Accelerator

An anticipated M.2 M-Key AI acceleration module from Radxa is expected to simplify Edge-AI and on-device ML workloads. With its octa-core Cortex-A55 processor that can clock in at 1.5GHz and a purported 24 TOPS of INT8 precision compute NPU, the Radxa AICore AX-M1 is built around the AXERA AX8850 SoC and imparts AI processing capabilities in an inconspicuous package. The Radxa AICore AX-M1’s embedded VPU supports H.264/H.265 HD video codec for an assortment of streaming and sophisticated video analysis scenarios, and it also offers 8K video streaming and computational image processing for exceptionally detailed video analysis.  

The onboard VPU of the Radxa AICore AX-M1 accelerator module supports 16-channel 1080p decoding, for use in computer vision and multi-stream video analytics. The system module is capable of processing computationally demanding AI models without resorting to external memory because of its 8GB of dedicated LPDDR4X memory. The AX-M1 accelerator is readily compatible with mainstream industrial and embedded motherboards due to the fact that it uses the standard M.2 M Key 2280 form factor. Edge Computing applications include Smart cameras, Security surveillance, Consumer Electronics, Autonomous Driving, and Smart mobility. 

Radxa Edge AI Accelerator front

Radxa AICore AX-M1 Specifications:

• SoC: Axera AX8850 
  • Integrated Octa-core processor ARM Cortex-A55, up to 1.5GHz 
• NPU:
  • Built-in acceleration for Transformer-based model inference 
  • 24TOPS@INT8 precise computing for neural network inference processing 
  • Supports Matrix Arithmetic Unit (MAU) and Intelligent Video Engine (IVE) 
• System Memory: 8GB LPDDR4x RAM 
• VPU: 
  • Supports H.264/H.265 8K@30fps encoding and decoding 
  • Supports decoding of 16 channels of 1080p@30fps 
  • Supports real-time encoding of 7680 x 4320@30fps + 1080p@30fps 
• Form Factor: Uses standard M.2 2280 M Key form factor 
• Dimensions: 22 mm x 80 mm 
• Operating Voltage: 3.3V 
• Power Consumption: ≤ 8W 

In terms of Hardware compatibility, the Radxa AICore AX-M1 M.2 accelerator module is compatible with a wide range of mainstream host system platforms, including x86-based platforms from Intel and AMD as well as Rockchip-based ARM systems. Talking about the software section, the Radxa AICore AX-M1 M.2 accelerator module is compatible with various Linux based operating systems including Ubuntu, Debian, and CentOS. 

Radxa AICore AX-M1 integrated with ROCK 5B+Board

Radxa has provided comprehensive documentation on their Wiki, including sections on hardware installation, driver setup, and model deployment. The AICore AX-M1 supports acceleration of a range of popular models including large language models such as DeepSeek-R1-Distill, Qwen2.5/3, Llama3.2, and Gemma2; vision models like MixFormerV2, Real-ESRGAN, InternVL3, CLIP, and the YOLO family; speech models including Whisper and MeloTTS; and text-to-image generation with Stable Diffusion v1.5. 

The Radxa AICore AX-M1 doesn’t appear to be available for purchase yet. More information on the AICore AX-M1 is available on the official product page. 

Images used courtesy of Radxa.