10.5281/zenodo.3550742
https://zenodo.org/records/3550742
oai:zenodo.org:3550742
Debdeep Paul
Debdeep Paul
Real-Time Server Monitoring and CNN Inference on FPGA
Zenodo
2019
CERN openlab
summer-student programme
2019-11-22
10.5281/zenodo.3550741
https://zenodo.org/communities/cernopenlab
Creative Commons Attribution 4.0 International
Neutrinos are subatomic particles, very similar to an electron, but without any electrical charge and
a very negligible rest mass. They are the most abundant and perhaps the most mysterious matter
particles in the universe!
Detecting and tracking neutrinos poses a different challenge both in terms of hardware and
algorithms. Since they have very little interaction with matter, they are incredibly difficult to detect.
So, every bit of information produced by the neutrino interactions is required to be stored and
analyzed at an incredible fast rate.
Present day State-of-the-Art Convolutional Neural Networks (CNNs) have shown an un-precedented
accuracy on many modern AI applications. They have become the de-facto standard for a wide range
of tasks ranging from computer vision to machine translation due to their high accuracy and
robustness.
Research has proved that CNNs also hold the potential to excel in High Energy Physics (HEP)
applications as compared to traditional methods in identifying particle interactions in sampling
calorimeters or Time Projection Chambers. However, these algorithms are both computationally
and memory extensive which limits them from running on normal CPU for real-time and power
constrained applications. Thus, customized hardware implementation of machine learning
algorithms can be a promising solution for higher performance and improved throughput.
Another important aspect is to constantly monitor the server which hosts the FPGAs for the CNN
inference and sent alert messages in crucial situations.
This project involved the development of a lightweight, easy to use solution for the server
monitoring so that all the important metrices can be visualized at a glance in real-time. Secondly, to
compare the CNN inference between the server and the FPGA in terms of accuracy, power
consumption and latency. The project also focuses on dealing with the numerical overflow problem
due to fixed point computations in FPGA during CNN inference.