Change Log

DX-RT v3.4.0 (July 2026)

Changed

• Debian package bundles dx_engine Python wheels for Python 3.8 – 3.14.
• CLI binary names updated for consistency: dxrt-cli → dxcli, parse_model → dxparse, run_model → dxrun
• Old names (dxrt-cli, parse_model, run_model) remain functional as backward-compatible aliases (symlinks on Linux/macOS, hard links on Windows)
• Python bindings now use dxrt_cxx_api.h (header-only) instead of direct C++ linkage
• libdxrt.so exports C symbols only (dxrt_*); internal C++ symbols hidden via version script
• Update minimum versions
  • Driver : 2.4.0 -> 2.5.0
  • PCIe Driver : 2.2.0 -> 2.4.0
  • Firmware : 2.5.2 -> 2.7.0
• Code style cleanup across Python package (inference_option.py, runtime_event_dispatcher.py, utils.py, etc.)
• Upgrade extern/include/dxrt/extern/cxxopts.hpp from v3.1.1 to v3.3.1
• Increase firmware update sequence wait time from 2s to 4s.
• Replace top-level matplotlib imports with lazy loading via ensure_dependencies() in plot.py to provide install guidance instead of ImportError

Fixed

• Improved shared-memory performance.
• Fix SEGV in NFHLayer caused by dereferencing uninitialized Request::_task in profiler instrumentation
• Return INT16_MIN instead of 0 from DeviceStatus::GetTemperature() when channel index is out of range
• Improve dxrtd error messages when manually executed
• Fix missing separator between devices and buffer_count fields in InferenceOption.__repr__ output
• Update ppcpu logic for hardness
• Fixed a critical crash (nullptr) in pyRunBenchmark by ensuring the memory lifetime of input buffers.
• Improve temperature range validation in dxtop.
• Fix multi-input dictionary handling issue.

Added

• Script to build dx_engine wheels per Python version.
• Documentation describing the Debian prebuilt directory structure.
• Script to generate the prebuilt/ directory.
• Stable C ABI (dxrt_c_api.h, 103 functions) enabling prebuilt SDK distribution without recompilation
• Header-only C++ wrapper (dxrt_cxx_api.h) for single-include modern C++ usage
• SDK-compat bridge headers (legacy/) for backward compatibility with existing dxrt_api.h code
• H1M firmware compatibility check: distinguish H1M (LPDDR4) from H1 (LPDDR5/LPDDR5X) in firmware update and device detection; support mixed 4-pack / 6-pack H1M configurations
• Deprecated Python API migration guide added to Python API Reference documentation
• Add shared memory-based inter-process communication using memfd
• Add dynamic IPC infrastructure with packet-based protocol layer
• Add cross-platform support for Linux and Windows IPC transport
• Add new device monitoring APIs (memory usage, per-core temperature, and utilization)
• Add a tool (plot_html.py) to generate HTML visualization reports for profiling data (profiler.json)
• Profiler
  • Add GetJobMetrics() API for comprehensive per-job profiling
  • Enhance performance statistics with Coefficient of Variation (CoV) metric
• Add input dtype validation to prevent undefined behavior caused by mismatching NumPy array types.
• Add libdxrt-bin Debian packaging for pre-built binaries with amd64/arm64 auto-detection.

---

DX-RT v3.3.2 (May 2026)

Changed

• Removed redundant build artifacts and temporary directories from the Debian package.

Fixed

• Improve Python extension module linking for _pydxrt build.

Added

• Added conditional pip upgrade (v21.3+) to ensure build stability on legacy OS environments.

---

DX-RT v3.3.1 (April 2026)

Changed

• Change the version of pre-built onnxruntime(1.23.2 -> 1.22.0) and openvino(25.4 -> 25.1)
• Enhance uninstall script to remove debian packages, legacy binaries, and test files from previous versions

Fixed

• Fix typos in the document
• Fix error in uninstall logic

Added

• Add libdxrt 3.3.1 debian package with updated build and install pipeline

---

DX-RT v3.3.0 (April 2026)

Changed

• Update minimum versions
  • Driver : 1.8.0 -> 2.4.0
  • PCIe Driver : 1.5.1 -> 2.2.0
  • Firmware : 2.4.0 -> 2.5.2
• Update the Python module version to match the C++ Runtime version.

Fixed

• Fix PPU data transfer error during multi-process execution in H1 and Multiple M1 M.2 environments.
• Fix input data lifecycle issues in the Python Runtime Module.
• Fix intermittent interrupt exceptions in IPC Message Queue.

Added

• Add dxtop for No Service Mode
• Add an example that loads an entire model file into a memory buffer and performs inference directly using this memory buffer.
• Add python InferenceEngine from numpy array
• Add acceleration features for CPU operations (Requires separate option configuration and build)

---

DX-RT v3.2.0 (December 2025)

Changed

• Optimize PCIe DMA sequence for better performance.
• Update OS requirements in installation guide for debian

Fixed

• Optimize device memory footprint of PPU models

Added

• Add RuntimeEventDispatcher class for C++
  • RuntimeEventDispatcher is a singleton class that provides centralized event dispatching and handling for runtime events from the DX-RT system, such as device errors, warnings, and notifications.
• Add Python wrapper for RuntimeEventDispatcher class -> dx_engine 1.1.4
• Add --profiler option to enable profiling mode in run_model.py
• Add --buffer-count option to configure inference buffer count in run_model.py
• Add build.sh options

  --use_service_on  Enable the use of the service in the build.
  --use_service_off Disable the use of the service in the build.
  --use_ort_on      Enable the use of the ORT component in the build.
  --use_ort_off     Disable the use of the ORT component in the build.

• Add __version__ import to the main dx_engine module namespace
• Implement per-instance configuration of Input and Output buffer counts for the Inference Engine.
• Enable direct loading of the .dxnn model format from a memory buffer within the Inference Engine.
• Add RuntimeEventDispatcher for centralized event handling and logging.

---

DX-RT v3.1.0 (November 2025)

Changed

• Update minimum versions

  - Driver: 1.7.1 → 1.8.0
  - PCIe Driver: 1.4.1 → 1.5.1
  - Firmware: 2.1.0 → 2.4.0

• Standardize command-line argument parsing across all examples

  - Use argparse (Python) and cxxopts (C++)
  - Unified argument format: -m/--model, -l/--loops, etc.

• Update SanityCheck.sh

  - Remove redundant libdxrt.so location check (validated via run_model execution)
  - Hide ONNX Runtime version when built with USE_ORT=OFF

• Enhance build and installation scripts

  - Use 'python3 -m pip' instead of 'pip' for better reliability
  - Add Python version compatibility check in build.sh
  - Improve logging with color-coded messages
  - Add uninstall.sh script for project cleanup

• Improve profiler functionality

  - Add Flush and GetPerformanceData functions
  - Group events by base name instead of individual job/request entries
  - Limit duration details to 30 values per group for cleaner output
  - Add memory usage tracking with high usage warnings

• Improve dxparse CLI tool with enhanced functionality
• Support dynamic shape output for tail CPU tasks
• Update C++ exception handling to translate exceptions into Python

Fixed

• Fix multiple bugs

  - Multi-tasking bugs related to CPU offloading and PPU output buffer
  - PPU model format and layout issues
  - Multi-output and multi-tail configuration bugs
  - Tensor mapping errors in non-ORT inference mode
  - Output tensor mapping and memory address configuration
  - run_model error when -f option and -l loop count exceeds 1024
  - Bounding issues on service

Added

• Add support for V8 DXNN file format with PPU support
• Add model voltage profiler tool (run_model_prof.py)

  - Requires firmware ≥ 2.2.0 and driver ≥ 1.7.1

• Add dxbenchmark CLI tool for performance comparison across multiple models

  - Default execution of 30 loops when neither loop nor time options are specified
  - Automatically creates result directory when result path is specified

• Add system requirements check in install.sh

  - Verify RAM ≥ 8GB
  - Verify architecture (x86_64 or aarch64)
  - Exit with error if requirements not met

• Add model file format version range validation

  - Error messages now indicate valid version range (min to max)

• Add time-based inference mode to run_model (-t, --time option)
• Add error handling for invalid firmware files and update conditions
• Add dxtop enhancements

  - Display PCIe bus number
  - Update cross-compilation script

• Add new Python APIs for device configuration and status retrieval
• Add new documentation files for Inference API, Multi-Input Inference, and Global Instance
• Add examples for asynchronous model inference with profiling capabilities

---

DX-RT v3.0.0 (September 2025)

• Update the .dxnn file format to version 7 (from v6).
• Update C++ exception handling to translate exceptions into Python for improved error handling.
• Update the Python v6_converter with enhanced functionality.
• Add support for models with multiple heads, tails, inputs, and outputs, including new C++ and Python APIs.
• Add a new internal C++ converter for v6 models.
• Add new Python APIs for handling device configuration and status retrieval.
• Fix several multi-tasking bugs related to CPU offloading buffer management and PPU output buffer mis-pointing.
• Fix a bug in the process of setting the PPU model format and layout.
• Fix a critical bug affecting models with multi-output and multi-tail configurations.
• Fix tensor mapping errors that occurred in non-ORT inference mode.
• Fix a warning message in get_output_tensors_info and a vector access bug in _npuModel.
• Fix an issue that prevented error messages from being displayed.
• Fix flaws in output tensor mapping and memory address configuration.
• Enhance OS and architecture checks in installation scripts
• Update documentation to reflect changes in supported CPU architecture and OS requirements.
• Enhance build and uninstall scripts with common utilities and improved logging
• Add PCIe bus number display for dxtop
• Add profiling data memory usage tracking with high usage warnings.
• Update user guide document
• Force-disabled with a warning instead of throwing a runtime exception in builds that don't support USE_ORT.
• Add time-based inference mode to run_model (-t, --time option)
• Profiler now groups events by base name (before) instead of showing individual job/request entries
• Limited duration details to 30 values per group for cleaner output
• Fix run_model error when -f option and -l loop count exceeds 1024
• Fix bounding option issue on service
• Add error handling for invalid firmware files and update conditions.
• Add a function to check Python version compatibility in build.sh.
• Add new documentation files for Inference API, Multi-Input Inference, and Global Instance.
• Add examples for asynchronous model inference with profiling capabilities in both C++ and Python.
• Update minimum versions

  - Driver : 1.5.0 -> 1.7.1
  - PCIe Driver : 1.4.0 -> 1.4.1
  - Firmware : 2.0.5 -> 2.1.0

• Fix kernel panic issue caused by wrong NPU channel number
• Update DeviceOutputWorker to use 4 threads for 4 DMA channels (3 channels to 4 channels)
• Improve error message readability in install, build scripts

  - Apply color to error messages
  - Reorder message output to display errors before help messages

• Update Python Package version (v1.1.1 -> v1.1.2)
• Modify run_async_model and run_async_model_output examples
• Modify build.sh (print python package install info)
• removed some unnecessary items from header files
• use Pyproject.toml instead of setup.py (now setup.py is not recommended)
• Fix some rapidjson issue from clients.
• Remove bad using namespace std from model.h (some programs need change)
• Add usb inference module (tcp/ip) (MACRO : DXRT_USB_NETWORK_DRIVER)
• Add options to SanityCheck.sh

  - Usage: sudo SanityCheck.sh [all(default) | dx_rt | dx_driver | help]

• The build compiler has been updated to version 14 for both USE_ORT=ON and USE_ORT=OFF configurations.
• Fix an issue where temporary files from the ONNX Runtime installation would accumulate.
• Fix a cross-compilation error related to the ncurses library for the dxtop utility.
• Add Sanity Check Features

  - Dependency version check.
  - Executable file check.

• Add APIs to the Configuration class for retrieving version information.
• PCIE details displayed on some device errors
• dxrt-cli --errorstat option added (this shows pcie detailed information)
• Modify run_model logging to include host info (Linux only).
• Add Python examples for configuration and device status.
• Add Python API for configuration and device status. (dx-engine-1.1.1)
• Add functionality to query the framework & driver versions in the Configuration class.
• Add weight checksum info for service
• Add ENABLE_SHOW_MODEL_INFO build option and configuration item
• Update code for compatibility with v3 environment
• Enhance UI for better clarity, enabled dynamic data rendering, and added visual graphs for NPU Memory usage.
• Fix: fix dx-rt build error caused by pybind11 incompatibility with Python 3.6.9 on Ubuntu 18.04

  - Support automatic installation of minimum required Python version (>= 3.8.2)  
  - Install Python 3.8.2 if the system Python version is not supported
  - On Ubuntu 18.04, install via source build; on Ubuntu 20.04+, use apt install
  - Added support in install.sh to optionally accept --python_version and --venv_path for installation
  - Added support in build.sh to accept and use --python_exec
  - Added support in build.sh to optionally accept --venv_path and activate the specified virtual environment

• The default build option for DX-RT has been changed from USE_ORT=OFF to USE_ORT=ON. If the inference engine option is not specified separately, use_ort will be enabled by default, activating the CPU task for .dxnn models.
• Add dxtop tool, a terminal-based monitoring tool for Linux environments. It provides real-time insights into NPU core utilization and DRAM usage per NPU device.

---

DX-RT v2.9.5 (May 2025)

• Added full support for Python run_model.
• Updated the run_model option and its description
• Improve the Python API

  - InferenceOption is now supported identically to the C++ API.  
     - set_devices(...) → devices = [0]  
     - set_bound_option(...) → bound_option = InferenceOption.BOUND_OPTION.NPU_ALL
     - set_use_ort(...) → use_ort = True
  - Callback functions registered via register_callback now accept user_arg of custom types. (removed .value)
     - user_arg.value → user_arg
  - run() now supports both single-input and batch-input modes, depending on the input format.

• Modify the build.sh script according to cmake options.

  - CMake option USE_ORT=ON, running build.sh --clean installs ONNX Runtime.  
  - CMake option USE_PYTHON=ON, running build.sh installs the Python package.  
  - CMake option USE_SERVICE=ON, running build.sh starts or restarts the service.  

• Add dxrt-cli -v to display minimum driver & compiler versions
• Addressed multithreading issues by implementing additional locks, improving stability under heavy load.
• Fix crash on multi-device environments with more than 2 H1 cards. (>=8 devices)
• Resolved data corruption errors that could occur in different scenarios, ensuring data integrity.
• Fix profiler bugs.
• Addressed issues identified by static analysis and other tools, enhancing code quality and reliability.
• Add --use_ort flag to the run_model.py example for ONNX Runtime.
• Add run batch function. (Python & C++)

  - batch inference with multiple inputs and multiple outputs. 

• Minimum model file versions

  - .dxnn file format version >= v6  
  - compiler version >= v1.15.2  

• Minimum Driver and Firmware versions

  - RT Driver Version >= v1.5.0  
  - PCIe Driver Version >= v1.4.0  
  - Firmware Version >= v2.0.5  

---

DX-RT v2.8.2 (April 2025)

• Modify Inference Engine to be used with 'with' statements, and update relevant examples.
• Add Python inference option interface with the following configurations
• NPU Device Selection / NPU Bound Option / ORT Usage Flag
• Display dxnn versions in dxparse (.dxnn file format version & compiler version)
• Added instructions on how to retrieve device status information
• Driver and Firmware versions

  - RT Driver >= v1.3.3  
  - Firmware >= v1.6.3

---