MORE | Spring 2024

Performance Evaluation of TPUs and FPGAs for Deep Neural Network Inference

The high computational complexity of AI algorithms and the stringent cost-energy-latency necessitate the use of efficient domain-specific hardware. With several design choices available from reconfigurable Field Programmable Gate Arrays (FPGAs) to highly specialized Application Specific Integrated Circuits for example Tensor Processing Unit (TPU), it is important to make a meaningful hardware platform choice. This research project aims to evaluate TPU and FPGA for inference by running DNN models on TPUs and an FPGAs and evaluate the following of the devices: Performance (Inferences/second), energy cost per inference, and a flexibility metric to quantize the ability to adapt to newer DNN models.