Leopold Grinberg
Research Staff Member, DataCentric High Performance Computing – IBM Research
Invited Talk: Unsupervised Training and Unified Global Address Spaces
Abstract:
In this talk we will focus on computational aspects and implementation of numerical algorithms for Unsupervised Local Machine Learning. Specifically, on the the use of the Unified Virtual Address Space (UVAS) within a single OS image. We will discuss the impact of UVAS from the algorithmic development and performance standpoints. In particular we will review the implementation of loading of batches of data for training, reduction of gradients from multiple compute elements, and model update implementations using the UVAS where multiple GPUs process data, which is distributed over the GPUs and CPUs memories.
(Joint work with Dmitry Krotov)
Biography:
Dr. Leopold Grinberg is a Research Staff Member in IBM’s Research Division. Leopold has over 14 years of experience in high performance and parallel computing. He oversees development and develops applications for IBM’s advanced computing architectures (IBM’s CORAL systems with hybrid Power9 CPU+ NVIDIA’s Volta GPUs). Dr. Grinberg duties include algorithm development and performance optimization for massively parallel solvers. He has expertise in CFD, Spectral Element, Particle Dynamics, seismic, transport solvers; C/C++, Fortran, CUDA, MPI, OpenMP4.5. Dr Grinberg is also an Adjunct Instructor in Applied Mathematics at Brown University, where he received his PhD in 2009.