Axelera Metis Compute Board pairs Rockchip RK3588 SoC with 214 TOPS Metis AI accelerator

Axelera Metis Computer Board is a Rockchip RK3588 mini-ITX motherboard equipped with a Metis AIPU (AI Processing Unit) capable of delivering up to 214 TOPS, 16GB LPDDR4 memory for the CPU, and 4GB LPDDR4x RAM for the AI accelerator.

I first came across the Axelera Metis AIPU in M.2 and PCIe card in 2023, and I was rather impressed with the advertised 214 TOPS of AI performance promised in this form factor and relatively affordable $149 price tag considering the price-to-performance ratio. At the time, it was still hard to source the chip and module due to limited quantities, and in 2024, it became easier to evaluate the solution with the Axelera Metis PCIe Arm AI evaluation kit based on a Firey ITX-3588J mini-ITX motherboard equipped with a 214 TOPS Metis AIPU PCIe card. The new Metis Compute Board is similar, but in a more compact form factor since the PCIe card is gone and the Metis AIPU is soldered directly on a mini-ITX motherboard.

Metis Compute Board specifications:

• SoC – Rockchip RK3588
  • CPU – Octa-core processor with four Arm Cortex-A76 cores @ up to 2.4 GHz, four Arm Cortex-A55 cores
  • GPU – Arm Mali-G610 MP4 quad-core GPU with support for OpenGL ES3.2, OpenCL 2.2, Vulkan 1.1
  • AI accelerator – 6 TOPS NPU
  • ISP – 48MP ISP
  • VPU
    • 8Kp60 video decoding, decoding up to 32x 1080p30 H.265/HEVC channels
    • 8Kp30 video encoding
• AI Processing Unit (AIPU) – Axelera quad-core Metis AI accelerator with digital-in-memory computing (D-IMC) delivering up to 214 TOPS, up to 3.200 FPS for ResNet-50; connected over PCIe Gen3 x4 with the RK3588 SoC
• System Memory
  • 16GB LPDDR4 for SoC
  • Up to 16 LPDDR4x for AIPU (4GB on the version currently sold)
• Storage
  • 64GB eMMC flash
  • 2x SATA 3.0 ports
  • MicroSD card slot
• Video Output
  • HDMI 2.0 port
  • DisplayPort via USB-C up to 4,096 x 2,160 @ 24 Hz
• Networking
  • 2x Gigabit Ethernet RJ45 ports
  • Optional WiFi/Bluetooth and 4G LTE via M.2 sockets
• USB
  • 4x USB 3.1 Gen1 Type-A ports
  • 1x USB 3.2 Type-C port with DisplayPort Alt. mode
  • 3x headers for 6x USB 2.0
• Expansion
  • 1x M.2 E-Key socket
  • 1x M.2 B-Key socket
  • 40-pin GPIO header with ADC, I2C, UART, SPI
• Power Supply – 12V DC input
• Dimensions – 17 x 17cm (mini-ITX form factor)
• Weight – TBD
• Temperature Range – -20 to +70°C

Axelera provides a Board Support Package (BSP) based on Yocto Linux, and the Axelera Voyager SDK to simplie AI model deployment and optimization using TensorFlow, PyTorch, or ONNX models. The SDK offers the following:

• Model Zoo with (among others)
  • Image classication  – ResNet-50, MobileNetv2
  • Object detection – MobileNetv2-SSD, Yolov5/v5/v6/v7/v8/v9
  • Segmentation – DeepLabv3, Yolov8-seg, U-net
  • Pose estimation – Yolov8-pose
• Compiler support for models from Pytorch and ONNX.
• Libraries including all pre- and post-processing required to run end-to-end pipelines: scaling; cropping; normalization; format conversion; non-maximal suppression (NMS), and more.
• YAML description le is used to automatically generate the AI pipelines, which can then be run as a plugin to GStreamer or within an inference server.
• Built-in tools to test the accuracy and performance of models running on the Metis AIPU.

If you want more details about performance, Axelera has a benchmarks page on its website running EfficientNet, MobileNet, and various Yolo models. It also includes a performance comparison between the Metis and the 26 TOPS Hailo-8 AI accelerator. I could not find a performance/watt comparison, however.

Axelera is taking pre-orders for the Metis Compute Board for 509 Euros. That’s a 15% discount over the regular 599 Euros price tag, and only valid during the pre-order period. Deliveries are scheduled to start in June 2025.

Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress