n recent years, two trends seem to shape the evolution of HPC: (i) the widening gap between data movement and data processing costs - both in terms of time and energy- and (ii) the end of the exclusive reign of IEEE 32-bit and 64-bit floating point arithmetics with the advent of AI lower precision requirements. It appears thus pertinent to revisit the legacy applications in order to reduce data movement by using lower precision arithmetic and/or compression at the expense of performing extra computations.
In this talk we explore the impact of such an approach on two representative applications in seismic imaging and digital rock physics (DRP). The first application consists in the reverse time migration, where we reduce expensive I/O operations using a novel compute-bound GPU-resident compression algorithm. Based on the Tucker decomposition used for tensor contractions, our compression algorithm exploits the data sparsity of the 3D domain seismic solution. The second application is a Self Organizing Map (SOM) algorithm, which is a critical phase of a time constrained DRP workflow. We present a GPU-accelerated mixed precision implementation that takes advantage of NVIDIA GPU Tensor Cores.
We report the impact on performance and numerical accuracy for both applications on latest NVIDIA hardware systems.
WATCH VIDEO
