wiki:Hardware/fDevices/gCUDA

Version 24 (modified by jkol, 6 years ago) ( diff )

CUDA

Certain specialized nodes are outfitted with NVIDIA CUDA capable GPUs or accelerators. The tables below list brief specifications of currently installed CUDA cards and which domain/node they are installed in. Currently, the nodes that have CUDA cards do not have any wireless radios and are meant for offloading computation via ethernet networking. The nodes are accessible via the Control and Data networks of their respective domain. See the individual domain pages for more detailed information.

Available Cards

Model Architecture # of CUDA Cores Memory (GB) Memory Bandwidth (GB/s) Single-Precision Performance Double-Precision Performance Compute Capability More Info
GeForce GT 640 GDDR5 Kepler 384 1 40.1 803 Gflops ~33 Gflops (unofficial) 3.5 1 2
GeForce GTX 750 Ti Maxwell 640 2 86.4 1306 Gflops ~47 Gflops (unofficial) 5.0 1 2 3
Quadro 2000D Fermi 192 1 41.6 480 Gflops ~40 Gflops (unofficial) 2.1 1 2 3
Quadro K5000 Kepler 1536 4 173 2150 Gflops ~90 Gflops (unofficial) 3.0 1 2 3
Tesla C2050 Fermi 448 3 144 1.03 Tflops 515 Gflops 2.0 1 2 3
Tesla K40 Kepler 2880 12 288 4.29 Tflops 1.43 Tflops 3.5 1 2
Tesla K80 Kepler 4992 24 480 8.73 Tflops 2.91 Tflops 3.7 1 2 3
Tesla P100 12GB Pascal 4992 12 549 9.3 Tflops 4.7 Tflops 6.0 1 2 3

Domains with CUDA Capabilities

Domain Node: CUDA Card(s)
SB1 node2-1: GeForce GTX 750 Ti
SB4 node2-1: GeForce GT 640 GDDR5
SB8 node2-1: Quadro K5000
SB9 node2-1: Quadro 2000D
Grid node21-1, node21-2, node21-4 to node21-7: Tesla K80
node21-3: 2x Tesla C2050
node21-9 to node21-32: Tesla P100
Note: See TracWiki for help on using the wiki.