The rise of the GPU

Posted on 10 Nov 2009 at 15:25

• 
• 
• 
• 
• 

 1 of 5

Gallery

Referred to as a desktop superPC, the FASTRA platform contains four GeForce 9800 GX2 graphics cards, each housing a pair of Nvidia G92 GPUs. Each GPU has 128 cores, so that's a total of 1,024 cores being brought to bear on the task. Reconstructing a full 3D volume consisting of a billion elements, a job that took several days on a cluster of four quad-core PCs can now be polished off within an hour using FASTRA. Costing just 4,000 euros (around £3,600) to build, this is truly a lot of processing bang for your buck.

It's common for technologies developed in academia to enter the mainstream. This is certainly true of personal supercomputing. In a departure from its roots in high-performance graphics, Nvidia is fuelling this revolution with its Tesla GPU computing cards.

In a move away from traditional thinking, the Tesla C1060 GPU Computing Processor is essentially a graphics card without any video-output circuitry. Instead, it has been optimised for high-performance computing. Based on the same T10 Series GPU that powers many graphics cards, it has a 240-core processor with a 1.3GHz clock speed, 4GB of memory and achieves a speed of 933 gigaflops (that's billions of floating point operations per second). To put this into context, Intel's fastest processor, the 3.2GHz Core i7, has four cores and manages 51.2 gigaflops. AMD has followed a similar strategy with its GPU-based range of FireStream products, which are described as stream processors. Top of the range is the FireStream 9270, which performs 1.2 teraflops (1,200 gigaflops) from its 800-core GPU.

Neither Nvidia nor AMD build ready-to-go systems based on their GPU-based computing cards. However, third parties have been quick to use these products in personal supercomputers. These desktop beasts employ GPUs working in tandem with CPUs. The multicore CPU performs those tasks that are inherently sequential in nature. It also distributes work to one or more many-cored GPUs. Today, research and development departments are major users of this technology with imaging and simulation applications in areas as diverse as astrophysics, medical imaging, molecular dynamics and financial risk analysis. Much of the software is custom written, although, as we'll see later, this is already starting to change with the prospect of GPU acceleration for the masses.

UK-based PC system builder Workstation Specialists (www.workstationspecialist.com) claims its WSX218 Tesla is the UK's only system to have four Nvidia C1060 cards. In addition to these GPU-based computing cards, it has two quad-core Core i7 processors clocked at up to 3.2GHz, a system memory capacity of 64GB DDR3 and up to 2TB of primary disk storage. In this top-of-the-line configuration, the system shows the massive potential of GPU accelerated computing by clocking up 3.73 teraflops from its 960 GPU cores. Perhaps the price might be something of an eye-opener at just over £12,000 for a fully configured system, but in terms of the cost per flop this is very cheap.

Nvidia estimates that to get the same level of performance using ordinary processors would require no fewer than 48 quad-core chips. Since the cheapest quad-core CPUs cost about £220, to match the performance of a system with four Tesla C1060s GPU cards using conventional CPUs would cost in the region of £100,000 for the processors alone. That doesn't include the cost of building 24 dual-processor high-performance workstations to house them.