Arbitor: A Numerically Accurate Hardware Emulation Tool for DNN Accelerators

Explore a conference talk introducing Arbitor, a hardware emulation tool designed for accurately evaluating deep neural network (DNN) accelerators and their impact on model accuracy. Discover how Arbitor leverages modern machine learning compilers to enable fast prototyping and numerically accurate emulation of common DNN optimizations on general-purpose GPUs. Learn about the extensive sensitivity study conducted using Arbitor, which examines the effects of low-precision arithmetic, approximate computing, and sparsity-aware processing on popular models like ResNet, Transformers, Recurrent-CNN, and GNNs. Gain insights into the tolerance levels of DNN models for lower precision arithmetic operations, the effectiveness of piece-wise approximation for complex non-linear operations, and the impact of structured sparsity on model accuracy. Presented by researchers from the University of Toronto and Vector Institute, this talk offers valuable knowledge for those interested in hardware accelerator design and optimization for DNN training workloads.