Machine Learning Group

Towards Edge Computing: A Software and Hardware Co-design Methodology for Scientific Applications (IO/CC)

From particle physics to cosmology, photon science, biology, medical and material science, to national security programs, huge amounts of raw data are produced, and scientific discoveries rely on extracting interesting events quickly and efficiently. Artificial intelligence (AI) technologies are being heavily investigated for particle-physic-related applications due to thier versatility and high-fidelity.

Currently, AI algorithms run on a hardware unit, such as graphic processing units (GPUs) or AI accelerators, where the experimental and simulation data are served offline. Such hardware units require massive amounts of computing power during training and deploying of a deep learning model. For example, the power-hungry and high-latency nature of GPUs does not allow to be close to a detector’s frontend electronics. Recent efforts in sparsifying deep learning algorithms; spiking neural networks (SNN); and power-efficient, custom integrated circuit design techniques have motivated Brookhaven Lab’s investigations into edge computing for scientific applications.

This work will employ a software and hardware co-design approach, resulting in an energy-efficient low-latency neuromorphic network that can be designed and fabricated via a conventional custom integrated circuit fabrication process. This will facilitate AI-enabled ASICs (application-specific integrated circuits) to be deployed on detector readout printed circuit boards in future detectors, such as the Electron-Ion Collider and nEXO.

Publications

Miryala, S., Mittal, S., Ren, Y., Carini, G., Deptuch, G., Fried, J., ... & Zohar, S. (2022). Waveform processing using neural network algorithms on the front-end electronics. Journal of Instrumentation, 17(01), C01039.

Miryala, S., Zaman, M. A., Mittal, S., Ren, Y., Deptuch, G., Carini, G., ... & Katkoori, S. (2022). Peak prediction using multi-layer perceptron (MLP) for edge computing ASICS targeting scientific applications. In 2022 23rd International Symposium on Quality Electronic Design (ISQED) (pp. 1-6). IEEE.