Scaling Performance of Barracuda Virtual Reactor on Azure

Barracuda Virtual Reactor was recently tested on Azure’s state-of-the-art “ND96asr_v4” virtual machines using 8 different models with sizes ranging from 25M to 55M computational particles. The study is a showcase for Barracuda’s scalability performance across the models. Smaller simulations which in our tests had less than 25M computational particles, achieved maximum speed-up when running in multi-GPU mode using two NVIDIA A100 Tensor Core GPUs. Larger simulations achieved maximum speed-up when using four GPUs. The presentation will share details on the Azure GPU computational resources used in addition to highlighting the recent results and the scaling capabilities of Barracuda Virtual Reactor code on Azure.