Investigation of heterogeneous platform for deep learning
We are be investigating a heterogeneous platform for deep learning in which GPUs are applied for training and FPGAs are initially used for inference, and later on will be adapted for training with customizable data types. Ultimately, we will be looking at how to utilize both GPUs and FPGAs for training and inference in a tightly coupled system.
Work in progress
Week of September 4th
- Go over UG1023 and UG1021 tutorials
Running tools on iridium:
sdx &
- make sure your workspace is in /home/NetID/project/workspace
- device is xilinx:kcu1500:4ddr-xpr:4.0
Copy examples to your own space:
cd ~/project cp -r /opt/Xilinx/SDx/2017.1/samples . cp -r /opt/Xilinx/SDx/2017.1/examples .
- you now have your own copy of samples and examples in ~/project
- up-to-date list of examples is in https://github.com/Xilinx/SDAccel_Examples
Week of September 11th
- Continue with UG1023 and UG1021 tutorials
Documentation
Tutorials: https://www.xilinx.com/html_docs/xilinx2017_2/sdaccel_doc/index.html
Examples: https://github.com/Xilinx/SDAccel_Examples
- UG1023 - SDAccel Environment User Guide (ver2017.2)
- UG1021 - SDAccel Environment Tutorial: Introduction (ver2017.2)
Development documenttion
- UG1207 - SDAccel Environment Optimization Guide (ver2017.2)
- UG1253 - SDx Pragma Reference Guide ( ver2017.2)
Hardware-related