Abstract

With recent developments in artificial intelligence (AI), there has been an ever increasing need for computation power in space. These AI models are traditionally implemented on GPUs. However, these units come at a high cost in terms of weight and power usage. Due to these limitations, many organizations are turning to the Versal ACAP device, a system on chip (SOC) with hundreds of AI engines. This option delivers more computation per watt than a traditional GPU. The desire to use Versal ACAP in space has created a need to perform radiation testing on the device. Extensive testing has been performed on the memory and programmable logic of the SOC. However, research on the AI engines is limited due to the complexity of such a test. This thesis proposes a new methodology for performing heavy ion testing on the Versal ACAP AI engines and uses this methodology to generate cross sections of the memories, processing pipelines and data transfer units of the cores. This methodology uses JTAG, UART, and PCIe protocols to achieve high data extraction rates. This test was designed to work at the NASA space radiation facility (NSRL) by synchronizing the test with the pulse mechanism of the beam. Two separate tests were performed at the NSRL facility using this new methodology. During this testing, the cross section for the AI memories ranged from 9.62��−14 – 1.26��−11 ��2/�� over a linear energy transfer range of 1.53 – 28.31 ��. ��2/�� . Further cross sections were generated for the scalar unit, DMA transfer unit, fixed vector unit, and floating vector unit. This test was performed on all 400 AI engine tiles of the VCK190 and all 304 machine learning tiles of the VEK280.

Degree

College and Department

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

Rights

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