Skip to content

Add VIPLite Execution Provider for Allwinner VIP9000 NPU (A733/T527 SoCs) #28244

@evgen-pervenenko

Description

@evgen-pervenenko

Description:

Several popular ARM SBCs now ship with the Allwinner A733 and T527 SoCs, which include a VeriSilicon VIP9000 NPU capable of 3 TOPS (INT8). These boards are increasingly used for self-hosted AI workloads (photo recognition, object detection, etc.).

Hardware:

  • Allwinner A733 (Orange Pi 4 Pro, Radxa Cubie A7S/A7A)
  • Allwinner T527 (Radxa Cubie A5E, Avaota A1)
  • NPU: VeriSilicon VIP9000, 3 TOPS, supports INT8/INT16/FP16/BF16

Current Linux driver status:

  • Kernel module: vipcore.ko (loaded, device at /dev/vipcore)
  • Userspace library: libVIPhal.so (VIPLite API)
  • SDK: Allwinner ACUITY Toolkit publicly available on GitLab (no NDA)
  • Model format: NBG (compiled via ACUITY from TensorFlow/PyTorch/ONNX)
  • Documentation: https://docs.radxa.com/en/cubie/a7s/app-dev/npu-dev/

Use case: Applications like Immich (photo management) use ONNX Runtime for ML inference on ARM servers. Currently only CPU is available on these boards. A VIPLite execution provider would enable NPU acceleration for face recognition, image classification, and smart search on low-power ARM hardware.

Growing ecosystem: These boards retail for $35-55 and are popular in the self-hosted community. Allwinner recently released full datasheets, user manuals, and Linux SDK publicly, signaling increased openness.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions