GTC Silicon Valley-2019: Performance of a Compressible DNS Code on the Latest GPU Architectures
GTC Silicon Valley-2019 ID:S9678:Performance of a Compressible DNS Code on the Latest GPU Architectures
Maruthi NaliganahalliHanumantharayappa(Jawaharlal Nehru Centre for Advanced Scientific Research)
We'll present performance results of an in-house compressible direct numerical simulation code on the latest GPUs. This GPU-Accelerated code, called ANUROOP, solves 3D Navier-Stokes equations using an unstructured finite volume framework. Following up on our GTC 2018 talk, we'll show weak and strong scaling results on the latest NVIDIA Tesla V100 GPUs, demonstrating that performance on the V100 is about twice as fast as on the NVIDIA Tesla P100. We'll discuss how the code scales linearly up to 25 nodes of the Prometheus cluster, which corresponds to 200 V100 GPUs. We will also present the Volta architecture-specific optimization techniques used in ANUROOP that further enhance performance.