SEARCH SESSIONS

Search All
 
Refine Results:
 
Year(s)

SOCIAL MEDIA

EMAIL SUBSCRIPTION

 
 

GTC On-Demand

Performance Optimization
Presentation
Media
Multi-GPU Programming Techniques in CUDA
Stephen Jones (NVIDIA)
Systems with multiple GPUs in a single node are almost universal in the cloud and high-performance computing worlds, and are increasingly common in power-user desktop systems such as NVIDIA''s DGX station. Effective use of these GPUs is critical to s ...Read More
Systems with multiple GPUs in a single node are almost universal in the cloud and high-performance computing worlds, and are increasingly common in power-user desktop systems such as NVIDIA''s DGX station. Effective use of these GPUs is critical to scaling programs, but developers have typically treated them as independent machines. Targeting multiple GPUs from a single process offers the potential for far greater performance, especially with the advent of NVLink which transforms the way that these GPUs can cooperate. We will cover a number of techniques and pitfalls for direct multi-GPU programming in CUDA, then look in depth at one novel method of using NVLink to scale some programs with minimal effort.  Back
 
Keywords:
Performance Optimization, Programming Languages, GTC Silicon Valley 2018 - ID S8670
Streaming: