Supercomputing PCs

Overview Over the last few years, the scene on the high-performance computer market has witnessed even more dramatic changes than the general computer market which has also been extremely turbulent. As today's graphics card chips have gigantic capacities and significantly exceed normal CPUs in terms of transistors and complexity, it makes sense to outsource intensive computing applications onto the GPU. Professional users in the fields of medicine, the oil industry or geology profit enormously from this option. The new supercomputing PCs are designed especially for these target groups. The graphics card as parallel computer > Why GPU computing? >> What solutions are available? >> Areas of application >>

The graphics card as parallel computer In the past, graphics card application was restricted to 2 or 3-dimensional images. A CPU, however, was used to calculate the displayed images. The increasing complexity of the calculations prompted calculations to be made on the graphics card itself. In the last few years, we have witnessed an enormous interest in utilising the immense capacity of parallel GPUs for other operations beyond the traditional 3D graphics. GPUs have developed far beyond simple pipeline implementation which is limited to fixed functions. They have become flexible, programmable and multiple parallel cores.

Their computing performance was increasingly employed to solve general numerical problems. Initially, extremely specialised hardware and creative implementations were required for these operations. However, modern interpretations already support abstract parallel computing models such as stream processing. This complemented the graphics pipeline and was used as the foundation for implementation on a GPU. With the Compute Unified Device Architecture (CUDA) from NVIDIA, new hardware and software possibilities are now available which make stream processing appear too restrictive.

Why GPU computing?

The modern, programmable and flexible GPU is one of the most powerful computing tools on this planet. Since 2000, each individual core on GPU data works with IEEE floating point precision just as standard CPUs (alias "real computers"). The pure floating point precision of a modern GPU is considerably larger and is growing faster than today's multi-core CPUs. A feature that has attracted a lot of interest in the IT community. It has even acquired its own new field of operation named GPGPU (General Purpose Processing on GPU) which reflects the demand

for utilising the entire GPU capacity for wider-scaled applications. It is no longer restricted to graphics but involves the pure computing power as required by professionals for calculating simulations and model calculations.

Today's TESLA 1070 System has over 960 processor cores and delivers up to four Teraflop computing performance in a highly dense 1U system.

Is this level of parallelisation and throughput imaginable on dual or quad-core CPUs? No, it's not!

What solutions are available?

There are various options available for scaling a GPU computing system.

The simpliest option is to integrate multiple graphics cards into one computer. Today's mainboards currently offer up to 4 PCI Express 16x slots. Two of these slots are usually equipped with full data transfer rate required for installing the graphics cards. A number of these computers can be connected to the network at a relatively low cost to build high-performance computing clusters.

CUDA Workstation 1000W: Up to two TESLA 1060 cards Up to two Teraflop computing performance Up to 480 cores Up to 8 GB RAM and 8 GB graphics memory Up to 4 TB hard drive memory

PCI Express switches also present another scalability option for the cascadable connection of multiple graphics cards on one single PCI Express bus. The Tesla S1070 System houses in one height unit 4 Tesla graphics cards each with 1 Teraflop single-precision computing performance and a total of 16 GB graphics memory. The cards can be accessed from outside via a shared PCI Express port using a PCI Express switch.

		Front-end system - CUDA Workstation 1000R: 4 U rackmount Up to two interface cards for the Tesla 1070 System Up to 8 GB RAM
		Tesla 1070 System: Up to four Teraflop computing performance in one highly dense 1U system Up to 960 cores 16 GB high-speed memory Delivers up to 408 GB/s memory bandwidth for extremely fast data transfer High-speed data transfer with PCI Express 2.0 Rail assembly with one single screw

Areas of application

GPU computing products enable scientists to execute larger algorithms. The products are intended for rendering, medical research and complex data processing as GPUs are considerably more effective for parallel data processing than standard processors. Geoscientists, molecular biology or medical institutes also rely on enormous computing power for simulations.

Oil companies can also conduct complex geographical and seismic analyses with GPU computing technology, for example. Weather simulations can also be accelerated by a factor of 50.