Hasso-Plattner-Institut
Prof. Dr. Tilmann Rabl
 

Clemens Lutz

Affiliation: NVIDIA
Title: Scaling GPU-Accelerated Data Analytics to Large Data Volumes

 

Abstract

GPUs excel at data-parallel processing, due to their high bandwidth on-board memory and specialized hardware architecture. This has been extensively researched in prior work. However, GPUs' Achilles heel has long been the data rate at which they can access large data volumes. In this seminar, I will show that the modern CPU+GPU architectures such as NVIDIA Grace-Blackwell overcome this hurdle. Grace-Blackwell, with its NVLink C2C interconnect and hardware decompress engine, delivers data to the GPUs at very high input rates. I will introduce a prototype GPU query engine built by NVIDIA. This engine showcases the benefits of modern CPU+GPU architectures for relational query processing. Finally, I will discuss using NVLink to perform index lookups on large data, thereby reducing the data transfer volume. Overall, modern CPU+GPU architectures are highly relevant for GPU-accelerated data analytics.

Short CV

Clemns Lutz works as a developer technology engineer at NVIDIA. He investigates how to make GPUs the processing platform of choice for database management systems and data analytics. His publications on this topic have been recognized with the GI DBIS dissertation award and best paper awards at the SIGMOD and BTW conferences. Prior to joining NVIDIA, Clemens earned his PhD at TU Berlin, and studied at ETH Zurich and Imperial College London.