Geniatech AIM-M-K AI Accelerator Brings 40 TOPS AI to M.2 Edge Systems

Geniatech AIM-M-K M.2 2280 AI accelerator

The landscape of Edge AI grows in prominence and the two core pillars for industrial applications are computational density and deployment compatibility. Geniatech’s new AIM-M-K AI acceleration module with NXP Ara240 NPU allows to move from standard computer vision tasks to the edge deployment of large language models (LLMs). The SSD has a standard M.2 2280 (M-key) form factor and connects via a PCIe Gen4 x4 interface, which means it can be added to existing x86 and Arm-based systems without the need to change the base hardware.

The AIM-M-K acceleration module supports running Transformer-based architectures on edge devices, overcoming the traditional memory limitations, with up to 16 GB LPDDR4 onboard memory. AIM-M-K is purpose-built for edge inference workloads, with up to 40 TOPS of INT8 performance for generative AI and computer vision applications.

Previously, we covered the Forlinx FAI-ARA240-M, another M.2-based AI accelerator powered by the NXP Ara240 processor. We also covered the Radxa AICore AX-M1 Edge AI accelerator, which features 8GB LPDDR4X RAM and 8K Ultra HD video support. Feel free to check them out.

M.2 AI Accelerator Top

Geniatech AIM-M-K Specifications:

• AI Accelerator: NXP Ara-240 NPU
• AI Performance: Up to 40 TOPS
• Memory: 16GB LPDDR4 onboard memory (8GB optional)
• Supported AI Frameworks: TensorFlow, PyTorch, ONNX
• AI Benchmark Performance
  • Stable Diffusion 1.4: 7 seconds per image
  • Llama2-7B: 12 output tokens per second
  • MobileNetV1 SSD: 974 IPS with 1.03 ms latency
• Host Platform Support: Compatible with x86 and Arm processors
• Interface:
  • M.2 M-Key
  • PCIe Gen4 x4 interface
• Security Features:
  • Secure Boot support
  • Root-of-trust processor
  • Encrypted interface protection
• Power Requirements
  • Operating Voltage: DC 3.3V
  • Power Consumption:
    • TDP: 12W
    • Typical: 3W
    • Full Load: 8W
• Thermal Solution: Integrated heatsink cooling
• Environmental Specifications:
  • Operating Temperature: 0°C to 50°C
  • Storage Temperature: -40°C to 85°C
  • Operating Humidity: 5% to 90%
  • Storage Humidity: 5% to 95%
• Module Size/Thickness: 22 × 80 mm M.2 2280 form factor, 0.9 mm PCB thickness

M.2 AI Accelerator Bottom

With capabilities of handling models such as ResNet50, YOLO based detection networks, and generative AI workloads (ex: Stable Diffusion and LLaMA 2), using local inference for reduced latency; the AIM-M-K has been specifically engineered for these tasks.  

For consumers, there is no retail availability; the AIM-M-K (M.2 Module) will retails for $645. For more information it is recommended to go through the product page. 

Images used courtesy of Geniatech.