Product Introduction

Product Overview

The DEEPX DX-M1 M.2 module brings server-grade AI inference directly to edge devices. The DX-M1 delivers 25 TOPS of performance with only 2W to 5W of power, offering 20 times the performance efficiency (FPS/W) of a GPGPU while maintaining GPU-level AI accuracy.

_images/dx-m1.png

Detailed specifications

_images/dx-m1-size.png

Name Parameters
AI Computing Power 25 TOPS
Form Factor M.2 M key
Dimensions 22 × 80 mm
Interface PCIe Gen 3 ×4
Memory 4GB LPDDR5 + 1Gbit QSPI NAND Flash
Debug Interface UART0, JTAG1

Instructions for use

Install

Connect to the M.2 interface of the RK3588, turn on the power, and confirm whether the DX-M1 PCIe accelerator card can be recognized.

root@firefly:/home/firefly# lspci
0004:40:00.0 PCI bridge: Rockchip Electronics Co., Ltd Device 3588 (rev 01)
0004:41:00.0 Processing accelerators: Device 1ff4:0000 (rev 01)

Deployment Environment

  • Download code

git clone --recurse-submodules https://github.com/DEEPX-AI/dx-all-suite.git

  • Compile and install drivers

# Before compiling, you need to install Linux Headers on your device. Please refer to https://wiki.t-firefly.com/en/Firefly-Linux-Guide/first_use.html#linux-headers
cd /dx-all-suite/dx-runtime/dx_rt_npu_linux_driver/modules/
./build.sh -d m1
./build.sh -d m1 -c install

# After installation, you can see dxrt_driver using lsmod.
lsmod

  • Install dx_rt

cd ./dx-all-suite/dx-runtime/dx_rt
./install.sh --all
./build.sh --install /usr/local
sudo cp ./service/dxrt.service /etc/systemd/system
sudo systemctl start dxrt.service
sudo systemctl enable dxrt.service
cd python_package
pip3 install .
reboot

# After installation, you can check the DX-M1 status using commands.
dxrt-cli -s

  • Upgrade Firmware

# The firmware on the DX-M1 may be incompatible with the current SDK. You can update the firmware corresponding to the SDK first.
cd ~/dx-all-suite/dx-runtime/dx_fw
dxrt-cli -u ./m1/latest/mdot2/fw.bin

  • Test

# Download the pre-compiled model from https://developer.deepx.ai/article/modelzoo/ ,This test uses the YoloV5S model

# run_model is benchmark tool.
run_model -m ./YoloV5S.dxnn -b -l 100 -v

More Information