Trustworthiness Evaluation of Deep Learning Accelerators Using UVM-based Verification with Error Injection

EasyChair Preprint 15357

Abstract

Testing the reliability and trustworthiness of high-performance computing (HPC) applications has made Deep Learning Accelerators (DLAs) verification critically important. In this paper, we introduce a hardware verification framework with an error injection methodology based on the Universal Verification Methodology (UVM) for DLAs that is scalable, reusable, and efficient to test the robustness and resilience of Deep Neural Networks (DNNs) running on various DLA designs. Furthermore, the error injection methodology is applicable to simulation and hardware-assisted verification (HAV) platforms for emulation and FPGA prototyping. Our proposed error injection mechanism is evaluated using Nvidia Deep Learning Accelerator (NVDLA), an open-source DLA core.

Keyphrases: CNN, Deep Learning Accelerators, Error Injection, NVDLA, UVM, verification

@booklet{EasyChair:15357,
  author    = {Randa Aboudeif and Tasneem A. Awaad and Mohamed Abdelsalam and Yehea Ismail},
  title     = {Trustworthiness Evaluation of Deep Learning Accelerators Using UVM-based Verification with Error Injection},
  howpublished = {EasyChair Preprint 15357},
  year      = {EasyChair, 2024}}