The rise of the GPU

Posted on 10 Nov 2009 at 15:25

• 
• 
• 
• 
• 

 1 of 5

Gallery

Graphics cards also have high-capacity RAM allocations that permit access to large amounts of data to feed the many processing cores. This is an alternative approach to the smaller-capacity, higher-performance memory caches that are traditionally built into CPUs.

Another interesting difference concerns the use of multithreading, a feature that Intel first introduced under the name of Hyper-Threading Technology in some of the Pentium 4s, and has reintroduced in the recent Core i7 processors. A thread is an independent stream of instructions, and multithreading permits more than one thread to be run on a single core.

Generally, a core can't process more than one thread at the same time because it has only a single set of execution units. However, a core that's equipped for multithreading duplicates the circuitry that stores the current state of the core, so it can be switched between threads. The benefit comes mostly when the core is stalled, perhaps because it has incorrectly predicted a branch or because it needs data that isn't in its cache, and has to wait for data or code to be retrieved from the memory. Under these circumstances, a switch to another thread could permit processor clock cycles to be used that would otherwise be wasted.

Intel processors have supported a maximum of two threads per core and, despite requiring many clock cycles to switch between threads, have boasted a 15-30 per cent performance boost. Nvidia GPUs, on the other hand, support vastly more threads than that. Between its 240 cores, for example, the Tesla T10 GPU can support 32,720 threads and can typically switch between them on every clock cycle.

The Personal Supercomputer

In the realm of CPUs, the move to multiple cores has progressed only to a moderate extent because there's a limit to how many cores can be used effectively in general-purpose applications. The GPU has moved much further down that path because graphics processing is a lot easier to split into many parallel threads. Each thread can work on a different part of the screen and is largely independent of the others.

Graphics processing isn't unique in the degree to which a task can be split into many threads that are worked on in parallel. Many of the power-hungry applications that require the vast computing resources of a supercomputer also fall into this category. Consequently, GPUs are now starting to enter the upper echelons of superpowered computing.

An exciting development that was announced by the University of Antwerp in Belgium in May 2008 shows how this approach can provide a vast amount of computing muscle for a very modest outlay. The university's Vision Lab research group is involved in the development of new computing methods for tomography. Tomography is a medical technique that combines a large number of X-ray photographs, acquired over a range of angles, to create three-dimensional images of a patient's internal organs. Advanced processing methods can require several weeks of dedicated processing time on an ordinary PC.

Previously at the university, processing was carried out using a cluster of sometimes hundreds of PCs, but this approach is expensive, takes a lot of space, consumes a lot of power and requires considerable maintenance. The inherently parallel nature of these computations lends itself to using PC clusters, and for the same reason, it's an ideal task for GPUs, as the scientists at Antwerp proved.

The fastra platform