AMAIX In-Depth: A Generic Analytical Model for Deep Learning Accelerators

In recent years the growing popularity of Convolutional Neural Network(CNNs) has driven the development of specialized hardware, so called Deep Learning Accelerator (DLAs). The large market for DLAs and the huge amount of papers published on DLA design show that there is currently no one-size-fits-all solution. Depending on the given optimization goals such as power consumption or performance, there may be several optimal solutions for each scenario. A commonly used method for finding these solutions as early as possible in the design cycle, is the employment of analytical models which try to describe a design by simple yet insightful and sufficiently accurate formulas. The main contribution of this work is the generic Analytical Model for AI accelerators (AMAIX) for the estimation of CNN execution time on DLAs. It is based on the popular Roofline model. To show the validity of our approach, AMAIX was applied to the Nvidia Deep Learning Accelerator (NVDLA) as a case study using the AlexNet and LeNet CNNs as workloads. The resulting performance predictions were verified against an RTL emulation of the NVDLA using a Synopsys ZeBu Server-based hybrid prototype. By refining the model following a divide-and-conquer paradigm, AMAIX predicted the inference time of AlexNet and LeNet on the NVDLA with an accuracy 98%. Furthermore, this work shows how to use the obtained results for root-cause analysis and as a starting point for design space exploration.