Software overview
The software provided by DEEPX can be divided into two logical parts: a compile-time environment and a runtime environment. These two environments work together to enable efficient deployment and execution of neural network models on the DX-M1 NPU.
The compile-time environment typically runs on a host PC and is used to convert trained models, provided in ONNX format, into the optimized DXNN format required by the DX-M1. This process includes model optimization and compilation steps that prepare the network for execution on the target hardware.
The runtime environment runs on the target system, such as a SoM equipped with a DX-M1 accelerator. It includes the necessary drivers, libraries, and runtime services required to load and execute compiled DXNN models on the NPU, and to integrate AI inference into applications running on the system.
The following diagram provides a high-level overview of the two software stacks. The pale olive–colored boxes represent software components originally developed and provided by DEEPX and are described in more detail below.
Compile environment
| Component | Description | Resources |
|---|---|---|
| Model compiler (dx-com) | A specialized compiler that converts pre-trained ONNX models to hardware-optimized DXNN binaries that can run on DEEPX NPUs. | DEEPX Github |
| Model viewer (dx-tron) | A graph viewer for visualizing and analyzing DXNN model files. | DEEPX S/W Download (Login is required) |
Runtime environment
| Component | Description | Resources |
|---|---|---|
| Demo Apps (dx-app) | These are demo applications provided by DEEPX. These applications allow users to quickly become acquainted with the DX-M1 AI accelerator and to evaluate its functionality and performance. The demo applications can also be used as a template when developing custom applications that make use of the accelerator. | |
| Runtime library (dx-rt) | This is the core runtime software responsible for executing DXNN models on DEEPX NPU hardware. It includes APIs (C/C++ and Python), drivers, and firmware for seamless NPU operation. | |
| NPU driver (dx-rt-npu) | This is the software driver for the AI accelerator and provides the low-level interface to the hardware. While it is logically part of the runtime library (dx-rt), it is documented separately due to being maintained in a separate software repository. |
Version compatibility
The software components provided by DEEPX have version dependencies between them and must be used in compatible combinations. To help ensure correct integration, the DEEPX dx-all-suite repository includes a version compatibility table that specifies which component versions are intended to be used together. For detailed information, refer to docs/version_compatibility.md in the dx-all-suite repository.