Symptom

The sixfab-dx package is installed and dxrt-cli is on $PATH, but dxrt-cli -s prints no device found. lspci may or may not list the DEEPX entry.

Cause

The DEEPX kernel module (dx_dma) is not loaded into the running kernel. The Sixfab APT package builds the module via DKMS, so normally a reboot or kernel update triggers an automatic rebuild, but the load can fail if DKMS could not compile against the current kernel or if the system was not rebooted after install.

Fix

1. 1

   Confirm whether the kernel module is loaded.

   lsmod | grep -i dx

   You should see a dx_dma entry. Empty output means the module is not loaded. Proceed to step 2.

2. 2

   Reboot the Raspberry Pi 5.

   The DEEPX kernel module may not have been loaded into the current session, especially right after a fresh install. Reboot first, then re-check with lsmod | grep -i dx and dxrt-cli -s.

   sudo reboot
3. 3

   Reinstall sixfab-dx to rebuild the kernel module.

   Reinstalling forces APT to recompile the DEEPX kernel module via DKMS against the currently running kernel.

   sudo apt update sudo apt install --reinstall sixfab-dx
4. 4

   Re-check kernel messages for module load errors.

   dmesg | grep -i dx

   If lsmod still shows no DEEPX module after the reinstall, the kernel ring buffer is where the failure is logged. Look for DKMS build errors or missing kernel headers.

Symptom

Inference completes without errors, but throughput is well below the published figure for the model on this configuration.

Cause

Most often one of four things: the PCIe link trained at Gen 2 instead of Gen 3, the model is not actually running on the NPU (a raw .onnx falls back to CPU), a YOLO model was compiled without PPU so non-maximum suppression serialises on the CPU, or CPU-side capture and preprocessing are starving the NPU.

Fix

1. 1

   Confirm the PCIe link is at Gen 3.

   dxrt-cli -s

   The output should report PCIe : Gen3 X1. If it reports Gen2 X1, the host is not feeding the DX-M1 at full bandwidth: measured ceiling drops from 800–900 MB/s to 400–450 MB/s. Confirm dtparam=pciex1_gen=3 is present in /boot/firmware/config.txt, then reboot.

2. 2

   Confirm the model is running on the NPU.

   If you passed a raw .onnx file to the runtime instead of a compiled .dxnn, the runtime falls back to CPU inference. That fallback can be one to two orders of magnitude slower. Always compile to DXNN with DX-COM first. See Custom Models (DXNN SDK).

3. 3

   For YOLO models, recompile with PPU support.

   A YOLO model compiled without PPU (Post-Processing Unit) performs non-maximum suppression on the CPU after every NPU inference, which becomes the bottleneck at high frame rates. Recompile with the PPU flag enabled in DX-COM and the bounding-box output is returned directly from the NPU.

4. 4

   Watch NPU utilisation during inference.

   Run dxtop in a second terminal. Near 100 % NPU utilisation with low FPS means CPU-side preprocessing is the bottleneck (capture, color conversion, resize). Low NPU utilisation means the NPU is starved and is waiting on the CPU. Fix capture and preprocessing first. Both are the CPU's responsibility on this board, not the NPU's.

5. 5

   Use the async API.

   The dxrt API supports asynchronous inference. Submit the next frame to the NPU while it processes the current one. This hides preprocessing latency and significantly improves throughput.

Symptom

The Edge AI Expansion Board was working before a routine sudo apt upgrade or kernel update. After reboot, dxrt-cli -s reports no device found, lsmod | grep -i dx returns empty, and lspci may or may not still list the DEEPX entry.

Cause

Fix

Run the following sequence in order. The Q&A guidance is that a clean reinstall of sixfab-dx is sufficient to recover; the steps below add the verification checks needed to confirm the rebuild succeeded.

1. 1

   Confirm the kernel version actually changed.

   This rules out unrelated regressions. If the kernel version is unchanged, the OS update did not invalidate the module. Re-run the Quick diagnostics at the top of this page first.

   uname -r
2. 2

   Confirm the DEEPX kernel module is no longer loaded.

   lsmod | grep -i dx

   Empty output is consistent with the OS-update cause. DKMS did not produce a dx_dma module for the new kernel. Continue to the reinstall.

3. 3

   Refresh the package index, then reinstall sixfab-dx.

   The --reinstall flag forces APT (and therefore DKMS) to recompile the kernel module against the current running kernel even if the package version on disk has not changed.

   bash

   # 1. Refresh package index
   sudo apt update
   
   # 2. Reinstall sixfab-dx (recompiles the dx_dma kernel module via DKMS)
   sudo apt install --reinstall sixfab-dx

   Compilation against the new kernel is the longest part of the reinstall. Lots of compiler output is normal.

4. 4

   Reboot.

   A reboot ensures the freshly compiled module is loaded cleanly into the running kernel, rather than relying on the install-time hooks.

   sudo reboot
5. 5

   Verify the module loaded and the NPU is reachable.

   bash

   # Module loaded?
   lsmod | grep -i dx
   
   # Runtime + NPU healthy?
   dxrt-cli -s

   Expected output Ready

   DXRT v3.3.0
   =======================================================
    * Device 0: M1, Accelerator type
   ---------------------   Version   ---------------------
    * RT Driver version   : v2.4.0
    * PCIe Driver version : v2.2.0
   -------------------------------------------------------
    * FW version          : v2.5.6
   --------------------- Device Info ---------------------
    * Memory : LPDDR5 5600 Mbps, 3.92GiB
    * Board  : M.2, Rev 1.0
    * Chip Offset : 0
    * PCIe   : Gen3 X1 [01:00:00]
   NPU 0: voltage 750 mV, clock 1000 MHz, temperature 45'C
   NPU 1: voltage 750 mV, clock 1000 MHz, temperature 45'C
   NPU 2: voltage 750 mV, clock 1000 MHz, temperature 45'C
   =======================================================

   If the reinstall fails to compile

   If apt install --reinstall sixfab-dx reports DKMS build errors against the new kernel, the most common root cause is missing kernel headers. The Sixfab APT package normally pulls raspberrypi-kernel-headers as a dependency, but on a manually patched system it may be absent. If that fails, the safest path is a fresh setup against the current kernel. Follow the Quickstart install steps from the start.