Intel® FPGA AI Suite: SoC Design Example User Guide

ID 768979
Date 12/01/2023
Public

A newer version of this document is available. Customers should click here to go to the newest version.

Document Table of Contents

3.7.1. Running the M2M Mode Demonstration Application

The M2M dataflow model uses the dla_benchmark demonstration application. The S2M bitstream supports both the M2M dataflow model and the S2M dataflow model.

You must know the host name of the Intel® Arria® 10 SX SoC FPGA Development Kit. If you do not know the development kit host name, go back to Determining the Intel Arria 10 SX SoC FPGA Development Kit IP Address before continuing here.

To run inference on the Intel® Arria® 10 SX SoC FPGA Development Kit:
  1. Open an SSH connection to the Intel® Arria® 10 SX SoC FPGA Development Kit:
    1. Start a new terminal session
    2. Run the following command:
      build-host:$ ssh <devkit_hostname>

      Where <devkit_hostname> is the host name you determined in Determining the Intel Arria 10 SX SoC FPGA Development Kit IP Address.

      Continuing the example from Determining the Intel Arria 10 SX SoC FPGA Development Kit IP Address, the following command would open an SSH connection:
      build-host:$ ssh arria10-62747948036a.local
  2. In the SSH terminal, run the following commands:
    export compiled_model=~/resnet-50-tf/RN50_Performance_b1.bin
    export imgdir=~/resnet-50-tf/sample_images
    export archfile=~/resnet-50-tf/A10_Performance.arch
    cd ~/app
    export COREDLA_ROOT=/home/root/app
    ./dla_benchmark \
       -b=1 \
       -cm $compiled_model \
       -d=HETERO:FPGA,CPU \
       -i $imgdir \
       -niter=5 \
       -plugins_xml_file ./plugins.xml \
       -arch_file $archfile \
       -api=async \
       -groundtruth_loc $imgdir/TF_ground_truth.txt \
       -perf_est \
       -nireq=4 \
       -bgr
    
The dla_benchmark command generates output similar to the following example output for each step:
[Step 11/12] Dumping statistics report
count:             8 iterations
system duration:   174.3530 ms
IP duration:       112.1184 ms
latency:           79.9449 ms
system throughput: 45.8839 FPS
number of hardware instances: 1
number of network instances: 1
IP throughput per instance: 71.3531 FPS
IP throughput per fmax per instance: 0.3568 FPS/MHz
IP clock frequency: 200.0000 MHz
[Step 12/12] Dumping the output values
[ INFO ] Dumping result of Graph_0 to result.txt and result_tensor_boundaries.txt