Protons Evaluation of 7nm Versal™ AI Engine (AIE) Based Radiation Tolerant Platform for Deep Learning Applications

This paper presents the SEE evaluation of a platform and design approach for enabling radiation-tolerant deep learning acceleration, based on the 7nm Versal™ VC1902 AI Engine (AIE), for aerospace and space 2.0 proton environments. The presented platform is suitable for deep neural network (DNN) implementations emphasizing image classification and uses 396 AIE cores. It includes solutions to mitigate both radiation-induced Single Event Functional Interrupts (SEFIs) and network datapath corruptions. The radiation-tolerant deep learning platform combines AMD's AI engine (AIE), Deep Learning Processing Unit (DPU) IP, and AMD Soft Error Mitigation (XiISEM) to mitigate SEFIs. Furthermore, a technique known as Fault Aware Training (FAT) was applied to mitigate single-event effects (SEE) in the datapath effectively. Test results from a 64MeV proton beam experiment using the ResNet-50 Convolutional Neural Network (CNN) for image classification are presented. The Single Event Upset (SEU) rate, system-level SEFI rate, and neural network classification/datapath performance are compared between the standard non-mitigated and radiation-tolerant platforms. Compared to standard non-mitigated training techniques, the radiation-tolerant platform using fault-aware training methods dramatically improves overall system response. SEFI error rates are ~1.4 x 10–¹⁰ p/cm² at > 60 MeV and twice a year at 500km in LEO orbit. Datapath events are reduced by ~4X and events classification accuracy degradation > 5% are completely mitigated.