Pipelined approach to fused kernels for optimization of machine learning workloads on graphical processing units

US10223762B2 · kind B2 · utility

Assignee

• International Business Machines Corporation · US

Inventors

• Arash Ashari · Kirkland, US
• Matthias Boehm · Neuendettelsau, DE
• Keith W. Campbell · Ottawa, CA
• Alexandre Evfimievski · San Jose, US
• John D. Keenleyside · Pickering, CA
• Berthold Reinwald · San Jose, US
• Shirish Tatikonda · Santa Clara, US

Key dates

• Filing date: Mar 16, 2018
• Grant date: Mar 5, 2019
• Expiry date: Mar 16, 2038

Classification

• Technology area (CPC G): Physics
• CPC primary: G06N20/00
• WIPO field: Computer technology
• WIPO sector: Electrical engineering

Abstract

A method for optimization of machine learning (ML) workloads on a graphics processor unit (GPU). The method includes identifying a computation having a generic pattern commonly observed in ML processes. Hierarchical aggregation spanning a memory hierarchy of the GPU for processing is performed for the identified computation including maintaining partial output vector results in shared memory of the GPU. Hierarchical aggregation for vectors is performed including performing intra-block aggregation for multiple thread blocks of a partial output vector results on GPU global memory.