Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
SC13 GPU Technology Theater
The NVIDIA Co-Design Lab for Hybrid Multicore Computing at ETH Zurich Peter Messmer, HPC Devtech
Application
Computational Scientist’s Perspective
Application
Computational Scientist’s Perspective
Application
Computational Scientist’s Perspective
Hardware
Application
Computational Scientist’s Perspective
Hardware
Hardware
Hardware Designer’s Perspective
Hardware
Hardware Designer’s Perspective
Hardware
Application
Hardware Designer’s Perspective
Hardware
Application
Hardware Designer’s Perspective
Mismatch of Application and Hardware
Hardware
Application
Co-Design Process
Application
NVIDIA Co-Design Lab for Hybrid Multicore Computing at ETH Zurich
Foster collaboration between scientific user
community and NVIDIA
Get input from real-world hybrid applications
Support community in design decisions
Opened on November 1, 2013
Located on ETH Zurich campus
www.nvidiacodesignlab.ethz.ch
Why ETH Zurich?
Leadership in Hybrid Computing in Europe
Parent organization of CSCS
Large investment in hybrid software
HP2C
PASC
Localized application expertise
From HP2C to PASC
HP2C (2009 – 2013) Focus: Applications
PASC (2013 – 2016) Focus: Domain Networks
Material Science
Life Science Solid Earth Weather
Fluid Dynamics Cosmology
Another reason for ETH Zurich..
Piz Daint at CSCS, 5272 K20X GPUs
#6 in Top500, #1 in Europe
NVIDIA Co-Design Lab Near Term Activities
WP 1 Application Analysis and Optimization
WP 2 Future Architecture
Analysis
WP 3 Application
Visualization Environment
WP 4 Publication,
Outreach and Training
Example: CP2K
Material science application
Innovative algorithm: O(N3) -> O(N)
New possibilities
Core algorithm: sparse matrix multiply
Small block matrices
Requires efficient tiny matrix-matrix product
Different from batched dgemm
CP2K: Optimization of small batched DGEMM
Co-Design aspects
Input to application design
Input for future hardware
Possibly feed back into cuBLAS
batched 23x23x23
COSMO: Study on Tri-Diagonal Solvers
Small tri-diagonal solves for vertical
coupling
Optimal algorithm unclear
Parallel Cyclic Reduction? Thomas?
How does hardware balance affect
performance?
Outreach: Workshop Dec. 11, 2013
Presentations by existing and new
GPU users
Poll GPU activity at ETH/Zurich
www.nvidiacodesignlab.ethz.ch/news
Summary
NVIDIA Co-Design Center for Hybrid Multicore Computing at ETH
Zurich
www.nvidiacodesignlab.ethz.ch
Foster collaboration between NVIDIA and domain scientists
Provide input from real-world applications
Perform algorithm research for future generation systems
Open to everyone
Workshop on December 11, 2013