After clicking “Watch Now” you will be prompted to login or join.
A Mixed-Precision Machine Learning Approach to Accelerate Geostatistical Simulations and Prediction on GPUs
Sameh Abdulah, KAUST
GTC 2020
Geostatistics represents one of the most challenging classes of scientific applications due to the desire to incorporate an ever-increasing number of geospatial locations to accurately model and predict environmental phenomena. Geostatistics modeling involves solving systems of linear equations on a given covariance matrix, which makes the model training phase prohibitive at large scale. We'll present a mixed-precision numerical solver to accelerate Geostatistics applications while maintaining appropriate prediction accuracy during the inference phase. Our algorithm not only achieves an average 1.9x speedup on CPU/GPU heterogeneous systems, but also reduces the data movement cost. The algorithm can also benefit from half-precision computations using NVIDIA Tensor Cores to further boost performance. Prediction accuracy of the new mixed-precision model shows promising results on synthetic and real environmental datasets.