CUDA Graphs, October 13, 2021

Introduction

NVIDIA will present “CUDA Graphs” on Wednesday, October 13, 2021. This event is a continuation of the CUDA Training Series and will be presented by Matt Stack from NVIDIA.

Many HPC applications encounter strong scaling limits when using GPUs sooner than when using CPUs due to higher throughput. The latency associated with submitting work to the GPU can be a challenge to this strong scaling. CUDA graphs are a model for work submission in CUDA that helps improve this situation. A graph is a series of operations (such as kernel launches) connected by dependencies, which are defined separately from their execution. This allows a graph to be defined once and then launched repeatedly. Separating the definition of a graph from its execution enables a number of optimizations that reduce the cost of launching and executing work on the GPU. 

This presentation will demonstrate how to use CUDA graphs in an application and discuss in what circumstances an application speedup may be obtained by using them, followed by hands-on example exercises. Cori-GPU access will be provided for current NERSC users. Temporary OLCF Summit access will not be available for remote participants.

Date and Time: 10 am - 12 pm (Pacific time), Wednesday, October 13, 2021

The format of this event will be online only.

Registration

Registration is required for remote participation.  Please click the "Registration" drop-down on this page to register. 

Remote Connection Information 

Registration is required for remote participation.  Please click the "Remote Connection Details" drop-down on this page to register. 

Presentation Materials 

• Slides
• Recording: TBA (to be posted shortly after the event)
• Exercises: The example exercises for this module can be found in the "exercises/hw13" folder of this GitHub repo.