本文へスキップ

DXNN SDK v2.3.0 Released

· 約 2分
DEEPX Technical Writer
DEEPX Technical Writer
DEEPX Technical Writing Team

We are excited to announce DXNN SDK v2.3.0, a major release focused on runtime efficiency, security hardening, and unified application architecture. This release features significant optimizations in the YOLO post-processing pipeline and a complete architectural refresh of the DX-Stream module.

Highlights​

  • Faster YOLO post-processing: TopK-first filtering in DX-COM v2.3.0 reduces CPU overhead for DFL-based YOLO pipelines.
  • DX-Stream v3.0.0 refresh: Modular architecture with zero-copy (DMA-Buffer) support and no-root installation.
  • Unified DX-APP v3.1.0: 5-layer structure with C++/Python parity for consistent app development.
  • Model and tool expansion: 280+ models, 560+ examples, and new tools like dxtop, DxScale, and DxConvert.

Component Versions​

  • DX-Compiler v2.3.0 (DX-COM v2.3.0, DX-TRON v2.0.1)
  • DX-Runtime v2.3.0 (DX_FW v2.5.6, NPU Driver v2.4.1, DX-RT v3.3.0, DX-Stream v3.0.0, DX-APP v3.1.0)

Migration Notes​

  • DX-Stream v3.0.0: Move env vars from /etc/profile.d/ to ~/.bashrc and use dx-modelzoo for model management.
  • DX-APP v3.1.0: Update project layout to the new 5-layer structure.
  • Minimum dependency: DX-RT v3.3.0 requires NPU Driver v2.4.0+ and Firmware v2.5.2+.

Known Issues​

  • PReLU-based models may still show FPS degradation.
  • Some models (OSNet, RepVGGA2, YoloV9C) may show accuracy variability by host CPU and calibration dataset.

For full details, see DX-AllSuite Release Notes.