Axelera AI Unveils Metis M.2 AI Accelerator with PCIe Gen3 x4 Interface for Edge AI Applications

The Metis M.2 AI Edge Accelerator, developed by Axelera AI, is a high-performance AI inference module designed for constrained devices. It features a Metis AIPU (AI Processing Unit) with 1 GB DRAM, supporting PCIe Gen3 x4 for fast data transfer. With a peak INT8 performance of 214 TOPS and power consumption ranging from 3.5W to 9W, it efficiently runs AI models while maintaining low energy use.

It is ideal for security, robotics, Industry 4.0, medical diagnostics, surveillance, mobility, agritech, and logistics applications. This module handles AI tasks in real-time, like multi-camera security, gesture control in robots, smart farming, and medical scans. Both x86 and ARM architectures are supported, and leading computing platforms from brands like Dell, Lenovo, and Advantech are compatible. With support for multiple neural networks and AI pipelines, the processing unit uses quantization techniques to shrink model size while keeping accuracy.

The operating temperature ranges from -20 to +70°C, with optional active cooling. Security features like Secure Boot and Root of Trust ensure system integrity.

Axelera AI Metis M.2 Edge AI Accelerator Specifications:

• Host Interface: PCIe Gen3 x4 (4 GB/s bidirectional)
• AI Processing Unit (AIPU): 1x Metis AIPU
• AIPU Memory: 1 GB DRAM
• Peak Performance: 214 INT8 TOPS
• Thermal Solution: Optional active cooling
• AI Model Support:
  • Image classification (ResNet-50, MobileNet)
  • Object detection (YOLOv5/v6/v7/v8, SSD)
  • Segmentation (DeepLabv3, U-Net)
  • Pose estimation (YOLOv8-pose)
• Security Features: Secure Boot, Root of Trust
• Power Consumption: 3.5W – 9W (typical)
• Dimensions: 80 x 22mm
• Operating Temperature: -20°C to +70°C
• Form Factor: M.2 2280 M-key

Its software ecosystem powered by the Voyager SDK, makes software integration easy with tools for testing, improving, and deploying pre-trained models. Additionally, the SDK supports PyTorch and ONNX, letting developers import models while handling quantization and optimization automatically.

The SDK has a simple setup for easy deployment and also gives low-level controls for direct hardware access. Built-in AI models for things like image sorting, object spotting, segmentation, and pose tracking are included.

This module is technically different from similar types of products we have written previously, such as the Hailo M.2 Accelerator Cards, the RP2040 M.2 Card and others.

At the time of writing, the Metis M.2 card is up for pre-order at €179.00 ($195.46) and will start shipping in March 2025. You can find additional information on the product page along with a product manual.

Please follow and like us: