# OpenHLS: High-Level Synthesis for Low-Latency Deep Neural Networks for Experimental Science

**Authors:** Maksim Levental, Arham Khan, Ryan Chard, Kazutomo Yoshii, Kyle Chard, Ian Foster  
**arXiv:** [2302.06751](https://arxiv.org/abs/2302.06751)  
**Date:** 2023-02-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2302.06751v4/

## Key points

- Large datasets are common in scientific domains such as high-energy physics, materials science, and cosmology.
- Collected data must be compressed or filtered to reduce storage costs and meet latency requirements.
- Filtering mechanisms involve machine learning models or physics-based approaches that require efficient use of hardware platforms like FPGAs or ASICs.
- Deep neural networks have proven effective in filtering tasks but are not widely employed in data acquisition systems due to design and deployment complexities.
- OpenHLS is an open-source compiler framework that can translate high-level representations of deep neural networks into low-level representations suitable for deployment on near-sensor devices like FPGAs.
- OpenHLS has been evaluated on different workloads and implemented in a case study involving Bragg peak detection within the context of high-energy diffraction microscopy.
- Results show that OpenHLS can produce an implementation of the network with a throughput of 4.8 $\mu$s/sample - approximately four times faster than existing implementations.
- OpenHLS does not rely on any proprietary dependencies and is lightweight enough to facilitate easy deployment across various hardware platforms.
- OpenHLS provides a viable solution for translating deep neural networks into low-level representations suitable for real-time filtering applications in scientific domains with high data rates such as high-energy physics, materials science, and cosmology.

## Summary

The generation of large datasets is a common occurrence in many scientific domains, including high-energy physics, materials science, and cosmology. For instance, the LHCb detector at the Large Hadron Collider (LHC) generates approximately 2 TB/s of data while processing proton-proton collisions at 40 MHz. The Square Kilometer Array, a radio telescope that will be operational by 2027, is projected to sustain data rates exceeding 10 TB/s. To reduce storage costs and meet stringent latency requirements for filtering mechanisms, collected data must either be compressed or filtered to retain only the most relevant information during acquisition. Filtering mechanisms typically involve machine learning models or physics-based approaches that require efficient use of hardware platforms such as field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs). Deep neural networks have proven effective in various filtering tasks but are not widely employed in data acquisition systems due to design and deployment complexities. To address this challenge, researchers have developed OpenHLS - an open-source compiler framework based on high-level synthesis techniques that can translate high-level representations of deep neural networks into low-level representations suitable for deployment on near-sensor devices like FPGAs. The OpenHLS framework has been evaluated on different workloads and implemented in a case study involving Bragg peak detection within the context of high-energy diffraction microscopy. Results show that OpenHLS can produce an implementation of the network with a throughput of 4.8 $\mu$s/sample - approximately four times faster than existing implementations. Additionally, OpenHLS does not rely on any proprietary dependencies and is lightweight enough to facilitate easy deployment across various hardware platforms. In summary, OpenHLS provides a viable solution for translating deep neural networks into low-level representations suitable for real-time filtering applications in scientific domains with high data rates such as high-energy physics, materials science, and cosmology. Its ability to improve processing speed while maintaining accuracy makes it a valuable tool for researchers and scientists working in these fields.