Professor Mike Alles of Vanderbilt presented at the Vanderbilt Seminar Series.
Radiation effects intersect with artificial‑intelligence hardware because energetic particles perturb the electronic circuits that execute AI algorithms. Understanding how fault injection, single‑event upsets, parametric shifts, and total ionizing dose alter timing, speed, and functionality of processors and memories is essential for reliable AI deployment in space, defense, and other radiation‑rich environments. The rapid expansion of AI workloads amplifies the need to evaluate hardware resilience alongside performance and power considerations.
The discussion surveys the physical mechanisms by which radiation interacts with contemporary AI substrates, emphasizing conventional 7 nm CMOS (Intel 4) and emerging neuromorphic approaches that co‑locate storage and computation. It examines stacked three‑dimensional integrations of logic, carbon‑nanotube transistors, and resistive‑RAM (R‑RAM) devices, highlighting how analog weight encoding in R‑RAM can reduce memory bottlenecks but introduces new radiation‑sensitivity pathways. Comparative analysis of memory technologies—DRAM, SRAM, non‑volatile magnetic, spin, and resistive states—is presented, together with quantitative references to power‑draw projections that approach national‑scale consumption, underscoring the trade‑off between brute‑force silicon and low‑power neuromorphic designs.
Armed with this knowledge, learners can articulate how radiation degrades AI hardware performance and critically compare memory‑compute architectures in terms of radiation resilience. They are equipped to assess power‑budget implications for AI system design, especially when targeting constrained platforms such as satellites or lunar installations. Moreover, they can evaluate the relative merits of conventional CMOS versus neuromorphic implementations, identifying the weakest link in heterogeneous stacks and making informed engineering decisions for radiation‑hardened AI deployments.