Hardware

If you are interested in GPU programming you will need access to suitable GPU hardware. There are two options available either purchase your own or utilise the remote GPUs on the universities iceberg facility;

Using the GPU Hardware on Iceberg

The icebery facility has 8 Nvidia Tesla K40s and 8 Nvidia Tesla Fermi M2070s GPU units for GPU Computing. Using the Iceberg GPU facilities first requires an iceberg account. You can obtain one by following the instructions at;
You must then request access to the GPU nodes by emailing hpchub@sheffield.ac.uk and asking to join the GPU project group. Further information on GPU Computing on the iceberg cluster can be found below.

http://www.shef.ac.uk/wrgrid/gpu/iceberg

Buying your own GPU Hardware

There are two major GPU manufactures (NVIDIA and AMD) and a wide range of GPU hardware available depending on your computational requirements and budget. Broadly speaking NVIDIA has better software support and offers the CUDA programming language (with development tools such as NSight for code debugging). AMD cards support OpenCL (which is also supported by NVIDIA cards) and are typically seen as a more affordable option. The university has NVIDIA GPUs within the iceberg facility but supports OpenCL on its Intel range of multi-cores.

Within the NVIDIA range of GPUs there are three (excluding mobile) product lines which are as follows;

GeForce: This is the mainstream consumer graphics architecture optimised for games and supporting NVIDIA CUDA. It lacks some of the support of the Tesla range, has (slightly) slower memory bandwidths and reduced full precision performance. In general this is the most affordable option for GPU computing.

Tesla: The Tesla range is optimised for GPU Computing and supports advanced features such as improved full precision double support, faster PCIe communication, faster memory bandwidths, ECC memory correction etc. These cards are considerably more expensive than the GeForce range but are the standard for any serious (NVIDIA) GPU Computing in powerful workstations or servers.

Quadro: The Quadro range is aimed at the workstation market and has improved support for multiple displays and better performance for applications such as CAD and 3D model design. As with all three of the product lines, Quadro supports NVIDIA CUDA and OpenCL but is more expensive than GeForce.


The university provides details on their approved suppliers and manufactures of desktop PCS and laptops. You may find that the desktop machines available are not able to be configured to accept the GPU hardware you had in mind. It is possible to purchase from additional suppliers (assuming your desired PC can be classed as specialist hardware) or build your own machine from components. You will need to obtain permission from the procurement team. Please speak to your department finance officer to arrange this.

GPUs are quite power hungry and in some cases quite large. If you are intending to upgrading a PC with a new GPU you will need to make sure that your PCs power supply is able to meet the power requirements and that your case has enough physical space. CICS or your departments IT officer should be able to provide advice regarding this.

Hardware Donations from NVIDIA

NVIDIA provide academic Hardware donations as small "seeding" gifts intended to enable researchers to begin a new project and gain the preliminary results to support a larger proposal to other funding agencies. Their goal is "breadth-first" seeding; most donations are small equipment donations (generally one GPU). This is assessed through submission to their Academic Partnership Program.

Comments