Skip to main content
This section contains the problems frequently encountered during the installation and usage of the DX-M1 accelerator module, along with suggested solutions.
The module not being listed on the PCIe bus at all indicates that the problem is at the hardware level.Apply the steps below in order.
  1. Cut the power of the board and remove the module from its slot and insert it again. Make sure that the module is fully seated in the slot and that the retaining screw is tightened.
  2. Check that the power adapter you are using can supply the total consumption of the board and the module. Insufficient supply causes the module to fail to establish the PCIe link.
  3. Review whether there is an error related to the PCIe link in the kernel logs.
If the lsmod | grep dx command produces empty output, the driver has not been loaded into the kernel.First check whether the driver produces an error while it is being loaded.
You can try loading the driver manually.
The Permission denied error indicates that the user does not have permission to access the device file created by the driver.You can temporarily run the application as the privileged user.
For a permanent solution, you can inspect the owner and the permissions of the device file and add your user to the relevant group.
This error can be encountered in Python applications when the input buffer is created with np.zeros(). All of the virtual memory pages allocated with np.zeros() point to the same physical page, and the PCIe DMA driver produces an error when it sees the same physical page more than once.Create the input buffer in the following way.
The DX-RT runtime verifies the compilation version of the loaded model. The model having been compiled with a DX-COM version older than the runtime on the board causes this error.Check the installed versions.
To solve the problem, you need to recompile the model using a DX-COM version that is compatible with the runtime version.
The frame rate staying below expectations usually originates from the steps other than the NPU.First determine where the bottleneck is. You can use the benchmark tool to measure the pure performance of the model on the NPU.
If this measurement gives the expected values but your application runs slowly, the bottleneck is in the pre-processing, post-processing or display steps. In that case you can consider the following.
  • Use the asynchronous template instead of the synchronous one. This way the CPU can prepare the next frame while the NPU performs inference.
  • Disable visualization.
  • Use the C++ equivalents instead of post-processing steps written in Python.
If the measurement is low as well, check the temperature of the module.
If the temperature remains high, the hardware lowers its clock frequency in order to protect itself. In that case you need to improve the ventilation of the module or use a heatsink.
The compilation process being successful does not guarantee that the model will produce correct results. There are two most common reasons for this situation.
  • Pre-processing mismatch: The pre-processing steps defined in the configuration file during compilation must be exactly the same as the steps applied during training. Differences in the color space conversion, the normalization values and the resizing method lead to incorrect results.
  • Insufficiency of the calibration dataset: The images used during quantization not representing the environment in which the model will run causes the accuracy to drop noticeably.
To verify the model output, you can compare the result of the ONNX model on the computer with the NPU result using the same input.
If more than one process will use the NPU at the same time, the dxrt service that manages the resource sharing must be running.
If the service is not running, start it and enable it so that it starts automatically at boot.
There is a single M.2 2280 slot on the T3 Gemstone O1 development board. For this reason an NVMe SSD and the DX-M1 accelerator module cannot be used at the same time.If you need additional storage space, you can use the microSD card slot or a storage device connected over USB.

​
Support

If your problem is not covered in this section, you can request support through the T3 Gemstone Community Forum. Including the information below when you post makes it possible to solve the problem faster.