PasTiS: building an NVIDIA Pascal GPU simulator for embedded AI applications

We present PasTiS, a simulator for the NVIDIA Pascal GPU architecture family, with a focus on timing simulation. PasTiS supports a subset of the Pascal ISA, sufficient to simulate the execution of neural networks. We present this subset, as well as the underlying microarchitecture that we modelled using information available from NVIDIA, from scientific publications, and from our own experiments. We demonstrate the precision of the simulator by comparing it to measurements on the NVIDIA Jetson TX2 development board, on neural network applications.