Updates

Software and System Updates

Update (2018 May 1)

CUDA version 9.0.176 and 9.1.85 are now available on the GPU nodes. The cuDNN library versions 7.0.3 and 7.1.3 are available.

Nvidia driver updated to v390.46.

All GPU queues now have a 7 day runtime limit.

Update (2016 November 7)

GPU 0 in Kaiju3 is reporting infoROM corruption errors so has been disabled. Interactive sessions and jobs may still land on the Kaiju3 K20 node but will only use GPU 1.

Update (2015 October 16)

CUDA 7.5 now installed (default version via libs/cuda modulefile or explicitly via libs/cuda/7.5.18 modulefile). Previous versions of the toolkit are available via modulefiles. Applications such as Amber will use the version of the toolkit against which they were compiled.

Nvidia driver updated to v352.39.

The peer-to-peer GPU node, available via

qrsh -l k20duo_p2p bash

now fully support peer-to-peer CUDA applications. After loading the 7.5.18 modulefile, run simpleP2P to see confirmation.

The 4 x K20 nodes will be unavailable on 10-12 November 2015 for a CUDA course.

Update (2014 November 13)

New K20 nodes in production. There are now five K20 nodes (each node has two K20 GPUs) giving a total of ten K20 GPUs.

Update (2014 October 20)

The K20 nodes will be unavailable on 30th and 31st October 2014 for use by the IT Services Intro and Intermediate CUDA courses. Please ensure any sessions running on the nodes have finished before then.

Update (2014 June 11)

Two more hosts are on order, each with two more Nvidia K20 cards.