跳转到主要内容
SDK Version: 2.4.0

RELEASE_NOTES

DX-All-Suite v2.4.1 / 2026-07-29​

  • DX-Compiler: v2.4.1
    • DX-COM: v2.4.0
    • DX-TRON: v2.0.1 (Deprecated)
  • DX-Runtime: v2.4.1
    • DX_FW: v2.7.3
    • NPU Driver: v2.5.1
    • DX-RT: v3.4.1
    • DX-Stream: v3.1.1
    • DX-APP: v3.2.1

Here are the DX-All-Suite v2.4.1 Release Notes.

What's New?​

This release adds DX-Benchmark (Beta), polishes the DX AI Studio (Beta) introduced in v2.4.0, and adds documentation updates across the suite.

📊 New: Reproducible NPU performance benchmarks​

DX-Benchmark (dx-benchmark) is now available in Beta — one standardized procedure that measures YOLO26 performance on any Host PC + NPU combination, published together with results for six hardware environments.

Highlights

  • Two measurement tiers — Model-Level (run_model): Latency (single-core, sync) and Throughput (multi-core, async); E2E Pipeline (DX-Stream): Single-Stream FPS and Multi-Stream channel capacity.
  • Comparable by construction — automatic ONNX-Runtime ON/OFF comparison, thermal steady-state normalization with throttle detection, and a full environment fingerprint recorded per run.
  • Published dataset — 6 hardware environments (H1-Quattro, Intel N97, OrangePi 5+, ROCK 5B+, and Raspberry Pi 5 with M1 / M1M) measured on v2.3.3 and v2.4.0.
  • Interactive dashboard — self-contained HTML (no CDN) for cross-environment and cross-version comparison; the same dataset backs DX AI Studio's Benchmark view.
  • Performance analysis — EN/KOR documents covering NPU-bound vs host-bound behaviour, ORT-mode selection, and per-environment deployment guidance.

Beta notice: DX-Benchmark is a Beta release (tool v0.1.0, measurement protocol v1). The CLI and output schema may change.

Learn more — see dx-benchmark/README.md and the performance analysis.

DX AI Studio (Beta) improvements​

  • Model Zoo — browse the full model catalog with one-click download (and re-download), in six languages.
  • App demos — correct, clearer results for classification, pose, and video demos.
  • Stream — demos and the Pipeline Builder now play reliably in remote / SSH browsers.
  • Compiler — smoother agentic auto-compile and clearer quantization diagnosis.
  • Benchmark — refreshed to the latest benchmark dataset.

For detailed updated items, refer to each environment & module's Release Notes.

DX-All-Suite v2.4.0 / 2026-07-22​

  • DX-Compiler: v2.4.0
    • DX-COM: v2.4.0
    • DX-TRON: v2.0.1 (Deprecated)
  • DX-Runtime: v2.4.0
    • DX_FW: v2.7.3
    • NPU Driver: v2.5.1
    • DX-RT: v3.4.0
    • DX-Stream: v3.1.0
    • DX-APP: v3.2.0

Here are the DX-All-Suite v2.4.0 Release Notes.

What's New?​

This major release focuses on AI-Powered Development Workflow, Next-Generation Quantization, and Production-Ready Stability. The introduction of Agent-Driven Development and advanced quantization techniques make this our most accessible and powerful release to date.

✨ New: Build DEEPX NPU apps with natural language​

DEEPX Agent-Driven Development (dx-agent-dev) is now available in Beta.

Describe an app or model task in plain language, and an AI coding agent drives the DEEPX knowledge base from start to finish — brainstorm → plan → TDD → verify — taking you from ONNX/.pt model compilation all the way to on-device DX-M1 NPU deployment.

Highlights

  • Natural-language workflow — turn a plain-language prompt into a working, on-device NPU app.
  • Multi-agent support — works with Claude Code, Cursor, GitHub Copilot, OpenCode, and Codex.
  • End-to-end coverage — model compilation (ONNX/.pt → .dxnn), inference app generation, and DX-M1 NPU deployment.
  • Ultralytics ecosystem — purpose-built for the Ultralytics model ecosystem on DEEPX NPUs.
  • Reproducible showcases — every bundled showcase ships with its prompt, measured results, and full build transcript.

Beta notice: dx-agent-dev is currently in Beta. Behavior and APIs may change as the feature matures.

Learn more — see Agent-Driven Development docs and the showcase gallery.

🖥️ New: DX AI Studio — one browser workspace for the DEEPX NPU (Beta)​

DX AI Studio is now available in Beta — an all-in-one desktop web workspace that unifies eight DEEPX tools in a single browser, in six languages.

Highlights

  • Eight tools, one hub — DX App (NPU inference on image/video/camera/RTSP with live multi-stream), DX Stream (real-time GStreamer pipelines with live WebRTC playback), DX Model Zoo (345 models with in-browser demos and one-click .dxnn/ONNX downloads), DX Compiler (ONNX → .dxnn with a config wizard, quantization tuning/diagnosis, and agentic auto-compile), DX EdgeGuide (board recommendation from real benchmarks), DX Benchmark, DX Monitor (live NPU telemetry), and DX Agent Dev.
  • Zero dependencies, self-installing — pure Python standard library; ./launcher.sh installs, boots every tool, and opens the browser with no manual setup.
  • Runs without hardware — every tool degrades gracefully to sample/mock data, so the whole studio is browsable with no NPU, SDK, or models.
  • Six languages and a built-in AI assistant (multi-provider, with a fully-offline local option).

Beta notice: DX AI Studio is a Beta release (studio v0.1.0). Features and UI may change before general availability.

Learn more — see the DX AI Studio User Manual (dx-ai-studio/docs/).

🎯 Advanced Quantization & Model Quality​

  • Automated Q-PRO Configuration: DX-COM now automatically generates DXQ combinations for higher-accuracy quantization, removing manual tuning requirements.
  • Quantization-Aware Training (QAT): End-to-end QAT support directly through dx_com.compile(), enabling fine-grained accuracy optimization.
  • QXNN Resume Flow: Re-run quantization with different settings without recompiling, dramatically shortening iteration cycles.
  • Quantization Diagnosis HTML Report: Visualize per-layer quantization quality with actionable recommendations.

🚀 Expanded Model & Platform Support​

  • Massive Model Library: Support for 347 models across 22 AI task categories, including 5 new tasks (3D Object Detection, Keypoint Detection, Object Pose Estimation, Panoptic Driving Perception, Hand Detection).
  • Extended Platform Coverage: Python 3.13–3.14 support (DX-COM) and Ubuntu 26.04 validation.
  • Windows Ecosystem: Full Windows MSVC support for DX-Stream including build, test suite, and Python bindings.

Key Updates​

Performance & Efficiency

  • Compiler Optimization: Faster compilation with reduced memory usage, especially on large models (DX-COM).
  • Stable C ABI: Prebuilt SDK distribution without recompilation via dxrt_c_api.h (103 functions) and header-only C++ wrapper (dxrt_cxx_api.h) for single-include modern C++ usage (DX-RT).
  • CLI Modernization: Updated CLI binary names (dxrt-cli → dxcli, parse_model → dxparse, run_model → dxrun) with backward-compatible aliases (DX-RT).
  • IPC Infrastructure: Shared memory-based inter-process communication using memfd with packet-based protocol layer and cross-platform support for Linux/Windows (DX-RT).
  • Python Ecosystem: Debian package bundles dx_engine Python wheels for Python 3.8–3.14 (DX-RT).

Stability & Fixes

  • Firmware & Hardware Stability:

    • Reverted M1/M1M IC, M.2 module, and DX-H1 Quattro board products PCIe device id to 0x0000 (DX_FW).
    • Adjusted CPU reset delay (20ms → 200ms) to ensure stable PLL lock (DX_FW).
    • Updated OTP Revision for improved hardware identification (DX_FW).
    • Disabled Root Complex Tx Equalization Preset 10 to prevent PCIe link compliance/test loops during normal boot (DX_FW).
    • Changed BAR0 type from prefetchable to non-prefetchable on VNPU board type (DX_FW).
    • Fixed input queue clearing when all bound options are deleted (DX_FW).
    • Solved PCIe enumeration issue on RZ/G3E (DX_FW).
    • Rejected in-flight mailbox commands during FW reboot and hardened recovery sleep (NPU Driver).
    • Fixed device recovery issues after firmware updates; resolved module installation errors in certain hardware environments (NPU Driver).
    • Automatic recovery logic for critical runtime error scenarios (NPU Driver).

  • Compiler Robustness:

    • Fixed Q-PRO/DXQ quantization crashes and stability issues observed on real models (DX-COM).
    • Fixed Python API integer input handling that caused accuracy degradation (DX-COM).
    • Fixed compilation errors in models with Split, Concat, Reshape, Bilinear Resize, Clip, or odd spatial dimensions (DX-COM).
    • Fixed NumPy 2.4+ and onnxruntime ≥ 1.25.0 compatibility issues (DX-COM).

  • Runtime Reliability:

    • Fixed critical crash (nullptr) in pyRunBenchmark by ensuring memory lifetime of input buffers (DX-RT).
    • Fixed SEGV in NFHLayer caused by uninitialized profiler instrumentation (DX-RT).
    • Improved shared-memory performance and temperature validation in dxtop (DX-RT).
    • Multiple API refinements: error messages, output formatting, ppcpu logic, multi-input dictionary handling (DX-RT).

  • Stream Stability:

    • LATENCY reporting: All elements now correctly account for processing time, stabilizing synchronization and QoS behavior (DX-Stream).
    • FLUSH recovery and push thread lifecycle: Proper reset of queues/threads/state on FLUSH, eliminating hangs on seek/replay (DX-Stream).
    • Error handling: Replaced abort() with proper GST_ELEMENT_ERROR messages across all elements (DX-Stream).

  • Application Fixes:

    • Super-resolution now preserves input resolution via dynamic tile padding (DX-APP).
    • Fixed --save option not producing output files in C++ sync runners for multiple AI tasks (DX-APP).
    • Multiple post-processor fixes: DOPE, YOLOPv2, RetinaFace, Depth Anything V2 normalization (DX-APP).
    • Build fixes: pybind missing includes, Windows x64 library paths, async runner metrics (DX-APP).

New Features & Tools

  • Compiler & Quantization:

    • Interactive HTML graph viewer for model inspection with parameter shapes and CPU/NPU partition reasons (replaces DX-TRON) (DX-COM).
    • dx_com.pre_optimize() API for ONNX-level pre-processing transforms with built-in YOLO post-processing integration (detection/segmentation) (DX-COM).

  • Runtime & Monitoring:

    • H1M firmware compatibility check: distinguish H1M (LPDDR4) from H1 (LPDDR5/LPDDR5X); support mixed 4-pack/6-pack configurations (DX-RT).
    • Device monitoring APIs for memory usage, per-core temperature, and utilization (DX-RT).
    • Profiler enhancements: GetJobMetrics() API for comprehensive per-job profiling and Coefficient of Variation (CoV) metric (DX-RT).
    • HTML visualization tool (plot_html.py) for profiling data (DX-RT).
    • Debian packaging improvements: prebuilt directory structure and libdxrt-bin with amd64/arm64 auto-detection (DX-RT).

  • Application Framework:

    • Native C++ post-processing for model zoo (YOLO families, semantic seg, Face3D, embedding/classification/attribute, restoration, YOLO-PPU), replacing Python fallbacks (DX-APP).
    • Opt-in --fast-postprocess path for object detection and instance segmentation (DX-APP).
    • YOLO Customizing Guide documentation (DX-APP).
    • Windows Visual Studio solution package extraction workflow with automatic OpenCV/DXRT CMake dependency configuration (DX-APP).
    • Build enhancements: --demo-models download option, minimal/category-based builds, Windows build selection TUI, --all flag for full build and install (DX-APP).
    • Knowledge base (.deepx/): specialized agents, app-building/SWE skills, and multi-platform agent-instruction generation (CLAUDE/AGENTS/copilot/cursor) via dx-agent-gen (DX-APP).
    • 7 new post-processors + pybind bindings (43 → 50 classes), 3 C++ factory interfaces, 3 visualizers for new AI tasks (DX-APP).
    • Per-model Python examples (4 variants: sync/async/sync_cpp_postprocess/async_cpp_postprocess) and C++ examples (sync/async) (DX-APP).

  • Streaming Infrastructure:

    • Multi-Stream Domain: Introduced application/x-dxvideoraw domain caps for unified processing of streams with different resolutions/formats (DX-Stream).
    • Base class migration: dxinputselector/dxgather → GstAggregator for N:1 multiplexing; dxvnpuoverlay → GstBaseSink for proper render/preroll lifecycle (DX-Stream).
    • TransformKernelPool: Automatic src-format-based kernel selection with libyuv fallback for dxscale, dxconvert, dxpreprocess, dxmsgconv (DX-Stream).
    • InferBackend abstraction: Refactored dxinfer to use Put/Get async pattern with backend property (auto/dxrt/dxvnpu) (DX-Stream).
    • Windows MSVC build and runtime environment: Full support including dependency check, build, demo launcher, test suite, Python binding (pydxs) (DX-Stream).
    • Test Suite: 73 new test binaries under test/base/{element,metadata,pipeline}/ covering element contracts, domain boundaries, end-to-end pipelines (DX-Stream).
    • DxMsgConv: include-frame property for base64 JPEG frame encoding with Kafka/MQTT consumer display support (DX-Stream).

  • Docker & Deployment:

    • Red Hat family support for the dx-compiler container: build and run on Fedora (42–45), RHEL/UBI (9, 10), and CentOS Stream (stream9, stream10) via new --fedora_version / --rhel_version / --centos_version options. dx-runtime and dx-modelzoo remain Ubuntu/Debian only (Docker).
    • Ubuntu 26.04 base image support for all three containers (Docker).
    • Faster dx-runtime / dx-modelzoo image builds: dx_rt and the matching dx_engine Python wheel are installed from prebuilt Debian packages instead of being compiled from source (Docker).
    • DX-Tron on Red Hat family images is provided as the web variant (run_dxtron_web.sh) since the .deb/AppImage is not supported there (Docker).

Known Issues​

  • PReLU Degradation: Significant FPS degradation has been observed in models using PReLU as an activation function.

Deprecation Notices​

  • DX-TRON: Deprecated as of v2.4.0. Migrate to DX-COM's standalone HTML graph viewer for model visualization.
  • PPU Type 2: Legacy PPU type 2 post-processing mode is deprecated. Migrate to dx_com.pre_optimize() API with built-in YOLO post-processing passes.

Migration Guide​

  • DX-TRON Users: Replace DX-TRON workflow with the new HTML graph viewer generated by DX-COM.
  • PPU Type 2 Users: Update compilation workflows to use dx_com.pre_optimize() API for YOLO models instead of PPU type 2.
  • NumPy 2.4+ Users: DX-COM v2.4.0 ensures compatibility with NumPy 2.4+ and onnxruntime ≥ 1.25.0.

For detailed updated items, refer to each environment & module's Release Notes.

DX-All-Suite v2.3.3 / 2026-05-14​

  • DX-Compiler: v2.3.1
    • DX-COM: v2.3.0
    • DX-TRON: v2.0.1
  • DX-Runtime: v2.3.3
    • DX_FW: v2.5.6
    • NPU Driver: v2.4.1
    • DX-RT: v3.3.2
    • DX-Stream: v3.0.1
    • DX-APP: v3.1.1

Here are the DX-All-Suite v2.3.3 Release Notes.

What's New?​

This hotfix release resolves a missing Debian package file issue in dx_rt_npu_linux_driver. (v2.4.1, packaging fix only).

DX-All-Suite v2.3.2 / 2026-05-11​

  • DX-Compiler: v2.3.1
    • DX-COM: v2.3.0
    • DX-TRON: v2.0.1
  • DX-Runtime: v2.3.2
    • DX_FW: v2.5.6
    • NPU Driver: v2.4.1
    • DX-RT: v3.3.2
    • DX-Stream: v3.0.1
    • DX-APP: v3.1.1

Here are the DX-All-Suite v2.3.2 Release Notes.

What's New?​

This patch release focuses on Debian package and Python build improvements in DX-RT.

  • DX-RT Build Quality: Removed redundant build artifacts from the Debian package, improved Python extension module linking, and added conditional pip upgrade for legacy OS compatibility.

Key Updates​

Stability & Fixes

  • DX-RT: Improved Python extension module linking for _pydxrt build.

New Features & Tools

  • DX-RT: Removed redundant build artifacts and temporary directories from the Debian package.
  • DX-RT: Added conditional pip upgrade (v21.3+) to ensure build stability on legacy OS environments.

For detailed updated items, refer to each environment & module's Release Notes.

DX-All-Suite v2.3.1 / 2026-05-06​

  • DX-Compiler: v2.3.1
    • DX-COM: v2.3.0
    • DX-TRON: v2.0.1
  • DX-Runtime: v2.3.1
    • DX_FW: v2.5.6
    • NPU Driver: v2.4.1
    • DX-RT: v3.3.1
    • DX-Stream: v3.0.1
    • DX-APP: v3.1.1

Here are the DX-All-Suite v2.3.1 Release Notes.

What's New?​

This patch release focuses on stability improvements, documentation corrections, license compliance, and firmware stability across the DX-Compiler installer, DX-Runtime, and DX-Stream modules.

  • Installer Fix (DX-Compiler): uninstall.sh now properly removes all installed packages and extracted module directories, and supports a new --target option for selective uninstallation.
  • Runtime Compatibility (DX-RT): Updated pre-built library versions (onnxruntime, openvino) for improved compatibility.

Key Updates​

Stability & Fixes

  • DX-Compiler Installer: Fixed uninstall.sh not removing installed packages (dx_com via pip3 uninstall) and the dxtron Debian package (apt-get remove), and not deleting extracted dx_com/ and dx_tron/ directories.
  • DX-RT: Updated pre-built onnxruntime (1.23.2 → 1.22.0) and openvino (25.4 → 25.1) for improved compatibility.
  • DX-Stream: Fixed uninstall not removing apps build directories and pydxs build cache.

New Features & Tools

  • DX-Compiler Installer: Added --target=<dx_com|dx_tron|all> option to uninstall.sh (default: all), consistent with install.sh.
  • DX-APP & DX-Stream: Added license information for third-party models and datasets.

For detailed updated items, refer to each environment & module's Release Notes.

DX-All-Suite v2.3.0 / 2026-04-10​

  • 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

Here are the DX-All-Suite v2.3.0 Release Note.

What's New?​

This major release focuses on Runtime Efficiency, Security Hardening, and a Unified Application Architecture. Significant optimizations in the YOLO post-processing pipeline and a major overhaul of the DX-Stream architecture make this our most robust release to date.

  • TopK-Optimized YOLO Pipeline: DX-COM v2.3.0 introduces a high-efficiency post-processing pipeline for DFL-based YOLO models, applying TopK filtering before decoding to drastically reduce CPU overhead.
  • Modular & Sudo-less DX-Stream (v3.0.0): A complete architectural refresh of the streaming module, featuring a new kernel abstraction for hardware backends and a user-friendly installation that no longer requires root privileges.
  • Unified APP Framework: DX-APP v3.1.0 establishes a strict 5-layer design pattern, ensuring 1:1 parity between C++ and Python implementations for seamless cross-language development.
  • Enterprise Linux Support: DX-COM validation has expanded to include Fedora 42–45, Red Hat 9–10, and CentOS Stream 9–10.

Key Updates​

Performance & Efficiency

  • DFL Post-Processing: Optimized pipeline for YOLO models decoder reduces CPU workload by filtering candidates via TopK before bounding box decoding.
  • NPU Latency: Reduced inference latency for the majority of supported models through compiler optimizations.
  • Zero-Copy Streaming: DX-Stream now supports DMA-Buffer zero-copy in RGA and DXOSD, reducing memory bandwidth usage.
  • CPU Acceleration: Introduced optional build-time features for CPU-bound operation acceleration in the runtime.

Stability & Fixes

  • Hardware Compatibility: Fixed 64-bit PCIe TLP errors on Raspberry Pi 4 and resolved ARM64 IOMMU DMA coherency bugs.
  • Multi-Device Handling: Fixed PPU data transfer errors and IPC exceptions occurring in multi-process/H1/multi-M.2 environments.
  • Robustness: Implemented DMA abnormal recovery for improved system resilience and fixed PPU safety via 2-cell memory alignment.
  • Power Saving: Re-enabled run mode NPU and hardened devMode race.

New Features & Tools

  • Expanded AI Tasks: Added native support for Depth Estimation (FastDepth) and Image Restoration (DnCNN, Zero-DCE, ESPCN).
  • Massive Model Zoo: Integrated 280+ models across 17 categories with 560+ examples, supported by a new manifest-based auto-download system.
  • Development Tools: Added dxtop for No-Service Mode monitoring and new GStreamer elements (DxScale, DxConvert) for HW-accelerated scaling and color conversion.

Known Issues​

  • PReLU Activation: Significant FPS degradation persists in models using PReLU activation functions.
  • Accuracy Variability: Certain models (OSNet, RepVGGA2, YoloV9C) may show accuracy variability depending on the host CPU and calibration dataset.

Migration Guide​

  • DX-Stream v3.0.0 (Breaking Change): Users must move environment variables from /etc/profile.d/ to ~/.bashrc. Models are now managed via dx-modelzoo; run setup_sample_models.sh to update.
  • DX-APP v3.1.0: Project structures have changed to follow the 5-layer pattern. Refer to the new src/ directory for updated implementation standards.
  • Minimum Dependencies: DX-RT v3.3.0 now requires NPU Driver v2.4.0+ and Firmware v2.5.2+ at a minimum.

For detailed updated items, refer to each environment & module's Release Notes.

DX-All-Suite v2.2.2 / 2026-02-26​

  • DX-Runtime: v2.2.2
    • DX-APP: v3.0.2
    • DX-STREAM: v2.2.1
  • DX-Compiler v2.2.1
    • DX-COM: v2.2.1

Here are the DX-All-Suite v2.2.2 Release Notes.

What's New?​

This release enhances the development experience with GPU-accelerated quantization and improved Windows platform support.

  • GPU-Accelerated Quantization: DX-COM now supports GPU quantization via JSON configuration with automatic GPU detection, significantly speeding up the compilation process.
  • Enhanced Windows Support: Improved Windows build stability with automated vcpkg installation and fixed compatibility issues.

Key Updates​

Performance & Efficiency

  • GPU Quantization: Added quantization_device support in JSON configuration for CLI-based GPU quantization. Automatic GPU detection falls back to CPU when CUDA is unavailable.
  • Windows Build: Automated DLL copying (dxrt, vkpkg) to dx-app/bin directory and added vcpkg installation script for streamlined Windows builds.

Stability & Fixes

  • Fixed DXQ enhanced quantization option bugs in DX-COM.
  • Fixed PPU compilation bug in Python Wheel Package for Python 3.8, 3.9, and 3.10.
  • Fixed an issue where compilation proceeded without error when invalid model input names were specified.
  • Removed experimental filesystem includes and updated float literals in example cpp files to resolve Windows build errors.
  • Refactored apply_argmax to reduce nesting and fix gcovr warnings.

Known Issues

  • Significant FPS degradation has been observed in models using PReLU as an activation function.

DX-All-Suite v2.2.1 / 2026-02-06​

  • DX-Runtime: v2.2.1
    • DX-APP: v3.0.1

Here are the DX-All-Suite v2.2.1 Release Note.

What's New?​

This release focuses on enhancing the YOLO ecosystem with expanded model variant support and improved flexibility in post-processing.

  • Extended YOLO26 Coverage: Full support for YOLO26 variants including classification, pose estimation, segmentation, and oriented bounding box detection.

Key Updates​

Stability & Fixes

  • Fixed hardcoded attribute size limitation in YOLO post-processing that could cause issues with models having different output configurations.

New Features & Tools

  • Added yolov26 cls, yolo26 pose, yolo26 seg, yolo26 obb examples

DX-All-Suite v2.2.0 / 2026-01-16​

  • DX-Compiler: v2.2.0
    • DX-COM: v2.2.0
    • DX-TRON: v2.0.1
  • DX-Runtime: v2.2.0
    • DX_FW: v2.5.0
    • NPU Driver: v2.1.0
    • DX-RT: v3.2.0
    • DX-Stream: v2.2.0
    • DX-APP: v3.0.0

Here are the DX-All-Suite v2.2.0 Release Note.

What's New?​

This release introduces a Python-Centric Ecosystem and a Complete Example Overhaul, making development more intuitive and integrated.

  • Python-First Workflow: DX-COM is now available via pip (Wheel), and new Python bindings (pydxs, dx_postprocess) allow for seamless metadata and post-processing management directly in Python.
  • Major DX-APP Refactoring (v3.0.0): Legacy demos have been replaced with a modern, task-oriented example system. This includes built-in support for the latest YOLO generations (v26, v10/v11/v12).
  • Expanded Hardware Acceleration: PPU (Post-Processing Unit) support has been extended to the newest YOLO models, further offloading CPU tasks to the NPU.
  • Advanced Resource Management: The introduction of NPU QoS (Quality of Service) and improved asynchronous handling ensures stable performance in multi-tasking environments.

Key Updates​

Performance & Efficiency

  • Extended PPU Support: Hardware-accelerated post-processing now supports YOLO26, YOLOv8, v9, v10, v11, and v12.
  • PCIe DMA Optimization: Reduced CPU dependency and improved sequence efficiency for high-speed data transfer (requires DX-RT SDK v3.2.0+).
  • Inference Reporting: Updated inf_time to include both NPU and PPU runtimes for realistic performance profiling.
  • Resource Prioritization: Added QoS to the NPU Scheduler to manage execution priority effectively.
  • Memory Footprint: Further reduced device memory usage for models utilizing PPU.

Stability & Fixes

  • PCIe Stability: Added PERST# signal wait during initial boot stage to ensure reliable link establishment.
  • Soft Lockup Prevention: Added sleep/reschedule logic in polling to prevent system hangs during slow hardware ACKs.
  • Stream Stability: Resolved race conditions and segfaults in secondary inference modes with shared buffers.
  • Model Accuracy: Fixed a known accuracy degradation issue in the DeepLabV3PlusMobilenet-1 model.
  • Code Robustness: Implemented global try-catch handling and improved argument validation across the application layer.

New Features & Tools

  • Installation & Deployment:
    • DX-COM Wheel: Install the compiler via pip for automated ML pipelines and Jupyter environments.
    • DX-TRON Debian: Added .deb package support for Ubuntu 20.04/22.04/24.04/26.04.
  • Development Tools:
    • YOLO26 Support: Integration of the latest Ultralytics model optimized for edge deployment.
    • RuntimeEventDispatcher: A new centralized C++/Python singleton for handling system events, errors, and warnings.
    • pydxs: New Python binding for managing Stream metadata (DXFrameMeta, DXObjectMeta, etc.).
  • Engine Capabilities: Enabled direct .dxnn model loading from memory buffers and per-instance I/O buffer configuration.
  • Testing Infrastructure: Established a Pytest-based E2E test system for DX-APP, achieving over 93% code coverage.

Known Issues

  • PReLU Degradation: Significant FPS drops may occur in models using PReLU activation functions.
  • PPU Conversion Gap: DX-Compiler v2.2.0 does not yet support converting face/pose models to PPU format (requires v1.0.0 for these specific tasks).
  • Breaking Changes: DX-APP v3.0.0 is not backward compatible with v2.x legacy demos or JSON configuration files.

Migration Guide

  • Example Transition: Move from the demos/ directory to the new src/cpp_example/ and src/python_example/ structures.
  • Configuration: Replace legacy JSON config files with the new Command-Line Argument system in Python (e.g., for YOLO26 execution).
  • Environment: Update your Python environment using the provided requirements.txt to support the new dx_engine and pydxs modules.

For detailed updated items, refer to each environment & module's Release Notes.

DX-All-Suite v2.1.0 / 2025-11-28​

  • DX-Compiler: v2.1.0
    • DX-COM: v2.1.0
    • DX-TRON: v2.0.0
  • DX-Runtime: v2.1.0
    • DX_FW: v2.4.0
    • NPU Driver: v1.8.0
    • DX-RT: v3.1.0
    • DX-Stream: v2.1.0
    • DX-APP: v2.1.0

Here are the DX-All-Suite v2.1.0 Release Note.

What's New?​

This release marks a significant step forward with new features and major stability improvements across all core components.

  • PPU Acceleration Integrated: The Post-Processing Unit (PPU) is fully integrated into the compiler (DX-COM), runtime (DX-RT), and streaming (DX-Stream) layers. This allows the NPU to handle NMS/bounding box decoding for models like YOLO and SCRFD, drastically reducing CPU overhead.
  • Next-Gen Model Support: The entire stack now supports the new DXNN V8 file format and DXNNv8 PPU models (DX-RT and DX_FW), enabling the newest generation of AI applications.
  • Windows Ecosystem: Full support for Windows 10/11 has been added to the DX-APP layer, complete with automated build scripts, making cross-platform development easier.
  • Advanced Diagnostics & Profilers: New dedicated tools like the dxbenchmark, and GstShark integration provide comprehensive performance evaluation and optimization capabilities.

Key Updates​

Performance & Efficiency

  • PPU Integration (Full Stack): PPU functionality is reinstated in DX-COM and integrated into DX-Stream and DX-APP to offload post-processing tasks (NMS/decoding) from the CPU.
  • LPDDR Stability: DX_FW reduced the LPDDR Training Margin (0.7 -> 0.62) and added enhanced margin testing logic to boost system stability.
  • Runtime Performance: DX-Stream enhanced buffer processing via direct buffer manipulation and disabled synchronization in the video sink (secondary mode).
  • Optimization Tools: DX-COM added the --aggressive_partitioning option and optimization level control (--opt_level 1).
  • Asynchronous Processing: DX-RT implemented the Asynchronous NPU Format Handler (NFH) for non-blocking inference.

Stability & Fixes

  • Critical Multi-Model Fixes (DX-RT): Resolved a critical bug affecting models with multi-output and multi-tail configurations and fixed several multi-tasking and CPU offloading buffer management issues.
  • Pipeline Stability (DX-Stream): Fixed a critical event processing timing issue in dxinputselector that caused compositor pipeline freezes.
  • LPDDR/Boot Stability (DX_FW): Fixed LPDDR frequency display issues after CPU reset, resolved PRBS training fail judge logic, and improved PCIe link-up stability (including RPi5 warm boot).
  • Windows Fixes (DX-APP/DX-RT): Fixed Windows MSBuild warnings using explicit static_cast (DX-APP) and fixed Windows environment compile errors (DX-RT).
  • Compiler Flexibility (DX-COM): Removed restrictions on key operators: Split, Transpose, Reshape, Flatten, and Slice.

New Features & Tools

  • DXNN V8 Model Support: Added support for the V8 DXNN file format and DXNNv8 PPU models across the stack.
  • Windows Support (DX-APP): Added full Windows 10/11 environment support with an automated build script (build.bat).
  • Advanced Diagnostics & Monitoring:
    • DX-RT: Added DX-Fit tuning toolkit, dxbenchmark (performance comparison CLI), and model voltage profiler.
    • DX-Stream: Added GstShark integration for comprehensive pipeline performance analysis.
    • DX_FW: Added Secure Debug and Model Profiling mode.
  • PPU Data Types: DX-APP added support for three new PPU data types: BBOX, POSE, and FACE.
  • DX_COM: Added Partial Compilation support (--compile_input_nodes/--compile_output_nodes).

Known Issues (DX-APP / DX-COM)

  • Accuracy degradation observed in the DeepLabV3 Semantic Segmentation model.
  • DX-Compiler v2.1.0 does not yet support converting face detection and pose estimation models to PPU format.

For detailed updated items, refer to each environment & module's Release Notes.

DX-All-Suite v2.0.0 / 2025-09-08​

  • DX-Compiler: v2.0.0
    • DX-COM: v2.0.0
    • DX-TRON: v2.0.0
  • DX-Runtime: v2.0.0
    • DX_FW: v2.1.4
    • NPU Driver: v1.7.1
    • DX-RT: v3.0.0
    • DX-Stream: v2.0.0
    • DX-APP: v2.0.0

Here are the DX-All-Suite v2.0.0 Release Note.

What's New?​

This release marks a significant step forward with new features and major stability improvements.

  • Performance Boost: The new "stop & go" inference function and an increase in DMA channel threads improve processing speed, especially for large models.
  • Enhanced Stability: Critical bug fixes, including a kernel panic and a Python compatibility error, make the platform more reliable across different environments.
  • Powerful New Tools: The new dxtop monitoring tool provides real-time insights into NPU performance, while a USB inference module expands connectivity options.
  • Expanded Model Support: The compiler now supports new operators like ConvTranspose, and most notably, offers partial support for Vision Transformer (ViT) models. This opens up a wider range of AI applications.

Key Updates​

Performance & Efficiency

  • Implemented a new "stop & go" inference function that splits large tiles for better performance.
  • Increased the number of threads for the DeviceOutputWorker from 3 to 4.
  • YOLO post-processing logic was updated to use a RunAsync() + Wait() structure to ensure correct output order.
  • The default build option for DX-RT is now USE_ORT=ON, which enables the CPU task for .dxnn models by default. Add automatic handling of input dummy padding and output dummy slicing when USE_ORT=OFF (build-time or via InferenceOption).

Stability & Fixes

  • Resolved a kernel panic caused by an incorrect NPU channel number.
  • Fixed a build error on Ubuntu 18.04 related to Python 3.6.9 incompatibility by adding automatic installation support for a compatible Python version (3.8.2).
  • Corrected a QSPI read logic bug that could cause underflow.
  • Addressed a processing delay bug in dx-inputselector and fixed a bug in dx_rt that affected multi-tail models.
  • In DX-COM, PPU(Post-Processing Unit) is no longer supported, and there are no current plans to reinstate it.

New Features & Tools

  • Added a new USB inference module.
  • Introduced a new terminal-based monitoring tool called dxtop for real-time NPU usage insights.
  • A new dxrt-cli --errorstat option was added to display detailed PCIe error information.
  • Support for the Softmax, Slice, and ConvTranspose operators was enabled.
  • Partial support for Vision Transformer (ViT) models was added.
  • Implemented a new uninstall script (uninstall.sh) for project cleanup.
  • In DX-RT, add support for both .dxnn file formats: v6 (compiled with dx_com 1.40.2 or later) and v7 (compiled with dx_com 2.x.x).

For detailed updated items, refer to each environment & module's Release Notes.

DX-All-Suite v1.0.0 Initial Release / 2025-07-23​

We're excited to announce the initial release of DX-All-Suite (DX-AS) v1.0.0!

DX-AS is your new integrated environment, bringing together essential frameworks and tools to simplify AI model inference and compilation on DEEPX devices. While you can always install individual tools, DX-AS ensures optimal compatibility by aligning all tool versions for you.

What's Included?​

This initial release provides a comprehensive suite to get you started:

  • Integrated Environment: A unified platform for all your DEEPX AI development needs.
  • Optimal Compatibility: Pre-aligned versions of individual tools to guarantee seamless operation.

Key Documentation​

To help you hit the ground running, we've prepared detailed documentation:

  • Introduction: Get a comprehensive overview of DX-AS.
  • Installation Guide: Step-by-step instructions to set up your environment.
  • Getting Started: A quick guide to begin using DX-AS.
  • Version Compatibility: Information on supported versions and configurations.
  • FAQ: Answers to commonly asked questions.

You can find all these resources and more in the docs directory of the repository.

Last updated: