National Nanotechnology Infrastructure Network

National Nanotechnology Infrastructure Network

Serving Nanoscale Science, Engineering & Technology

NNIN Cuda Workshop

                            Dates: August 12-14, 2009 Location: Harvard University

 Graphical Processing Units (or GPUs) are dedicated graphics devices whose ancestors date to the 1980s and the graphical chips at the heart of Atari Game consoles and others. GPUs are optimized to perform linear algebra functions, such as matrix multiplication, with a high degree of parallelism. Recently it has been discovered that the highly parallel functionality of GPUs makes them potentially useful for accelerating scientific calculation and simulation. NVIDIA Corporation, a maker of GPUs and the developer of the CUDA language for GPU programming, has recently named Harvard University a CUDA Center of Excellence. The principal research computer for the Center is the Orgoglio cluster; a joint resource of NNIN/C and the NSF-funded Cyber Discovery Initiative at Harvard.

Workshop NNIN/C will hold a hands-on tutorial wo rkshop on CUDA programming for the Graphical Processing Unit at Harvard University from August 12 to 14, 2009. This workshop will be limited to approximately 20-30 researchers/students and will focus on the basic concepts and examples from the physical sciences of parallel programming for the GPU. The workshop will include lectures each morning from accomplished GPU experts followed by afternoon, hands-on sessions in the computational lab. Lecturers Confirmed lecturers include: Richard Edgar, IIC, Harvard; Alan Aspuru-Guzik, Chemistry, Harvard; Mike Clark, Physics, BU; Hanspeter Pfister, SEAS, Harvard; Mark Watson, Chemistry, Harvard; Michael Stopa, NNIN/C.  

Schedule   Wednesday, August 12, LISE 303

 8 am – Registration (coffee and pastries)

 8:45 – 9:00 am – Welcome

 9:00 – 10:30 am – Hanspeter Pfister, "General Purpose Computing using Graphics Hardware"

 10:30 – 10:45 am – break

 10:45 – 11:30 am – Michael Stopa, "Job submission on Orgoglio and addone"

 11:30 – 12:45 pm – Richard Edgar, "Matmult I, the Slow and Easy Way"  

 Wednesday, August 12, Science Center B-09

 2:00 – 3:00 pm – Michael Stopa, logging in, submitting jobs, addone

 3:00 – 4:00 pm -- Richard Edgar, simple Matmult  

 Thursday, August 13, LISE 303

 9:00 – 9:45 am – Leslie Vogt, "The CUBLAS"

 9:45 – 10:30 am – Mark Watson, "Issues of Precision and MGEMM"

10:30 –10:45 am – break

 10:45 – 11:45 am – Mike Clark, "Linear solvers and CUDA"  

 Thursday, August 13, Science Center B-09

 1:00 – 2:00 pm – Mark Watson, lab on Monte-Carlo estimation of pi

 2:00 – 2:45 pm – Richard Edgar, Matmult using shared memory

  2:45 – 3:00 pm – break

  3:00 – 4:00 pm – Leslie Vogt, CUBLAS/Chopper lab  

  Friday, August 14, LISE 303

  9:00 – 10:00 am – Alan Aspuru-Guzik, "QChem on the GPU"

  10:00 – 11:00 am – Richard Edgar, "The Murchinson Widefield Array"

  11:00 – 12:00 noon – Claudio Andreoni, "Optimizing Newtonian Dynamics"  

  Friday, August 14, Science Center B-09

  1:00 – 2:00 pm – Won Ki Jeong, Image Processing lab

  2:00 – 3:00 pm – Claudio Andreoni, Newtonian dynamics optimization lab 

Orgoglio consists of 24 Nodes, each with: Single quad-core Xeon "Harpertown" processors at 3 GHz ·   16 GB of EEC DDR2 800 RAM ·   Two Tesla C1060 GPUs (each with 4GB of RAM) (total of 24 nodes/motherboards, 96 cores, 192 GB RAM, 48 S1070 cards). QLogic 24-Port 9024 DDR InfiniBand networking between the nodes.