Dirac: GPU Computing

Dirac Status: Up

Dirac is a testbed GPU cluster funded in collaboration with the Computational Research Division at Berkeley Lab, using funding from the DOE/ASCR Computer Science Research Testbeds program (DOE Contract Number DE-AC02-05CH11231). 

Dirac has 50 nodes of NVIDIA Tesla Graphics Processing Units (GPUs).   Four of these nodes consist of  one Tesla C 1060 GPU with 8 cores and 24 GB of memory, 44 consist of one Tesla C 2050 (Fermi) GPU with 8 cores and 24 GB of memory, one consists of 4 Tesla C 1060 GPU's, each with 8 cores and 24 GB of memory, and one consists of 4 Tesla C 2050 (Fermi) GPU's, each with 8 cores and 24 GB of memory

The purposes of the Dirac cluster are to investigate the applicability of GPUs to scientific calculation, gain experience with scheduling heterogeneous elements (such as GPUs) in a multi-user environment, and investigate programming models (e.g. MPI + CUDA, OpenMP + CUDA, etc.) for multi-GPUs to test scalability of code.