Abstract

With the recent increase of interest into artificial intelligence (AI), graphics processing units (GPUs) have been used as the main hardware architecture for AI computation. Because of their high power consumption and large size, it is necessary to consider other architectures to perform AI computing in outer-space where size, weight, power, and cost (SWaP-C) are tightly constrained. One such architecture is the Advanced Micro Devices (AMD) Versal Adaptive Compute Acceleration Platform (ACAP), which can achieve acceptable AI throughput at a lower power and size compared to a GPU. As the Versal ACAP gets used in space applications, it is necessary to test its performance and reliability in a harsh radiation environment. In this work we use the Versal ACAP to perform inference on an automatic modulation classification (AMC) model and test its reliability in the presence of radiation and soft errors. AMC is a classification technique that is used to identify the modulation scheme of a received signal without prior information. This can be used in a variety of applications such as threat detection, adaptable communication systems, and signal transmission security. While AMC models have achieved high accuracy, their deployment on space-grade hardware remains underexplored. We evaluate the AMC model accuracy, throughput, and power consumption using the Vitis-AI Deep Learning Processing Unit (DPU), which is a programmable engine for implementing deep learning designs on AMD hardware. We evaluate the model’s reliability on the DPU in the presence of soft errors by injecting faults into the programmable logic (PL) of the Versal during model inference. Of all the faults injected, we observe failures 8.05% of the time. We also perform an experiment running the model under a proton radiation beam and observed detected unrecoverable error (DUE) to have a cross-section of 9.62 × 10^(−11) cm^2 and silent data corruption (SDC) to have a cross-section of 1.85 × 10^(−9) cm^2. We also develop a custom hardware-implemented AMC model on the Versal and test it under fault injection.

Degree

College and Department

Ira A. Fulton College of Engineering; Electrical and Computer Engineering

Rights

https://lib.byu.edu/about/copyright/