GPGPU: General Purpose Computing on Graphics Processing Units

The graphics card is evolving backwards. Find out why and how it's outgrowing its specialist role.

If a chronicle on CPU and GPU architectures were to be written, the title ‘The decline of the CPU’ would not be at all unfitting. It is a miserable story really. In 1997, NVIDIA’s first successful graphics card, the RIVA 128 had 3.5 million transistors -– less than half of the 7.5 million transistors used to build the flagship CPU of the time, the Pentium II. Two years later, the tables had turned; NVIDIA’s GeForce 256 sported 23 million transistors, more than doubling the resources of Intel’s newest processor, the Pentium III. From there on, the GPU dominated the CPU in transistor resources. By 2004, Intel was no longer the favoured CPU for gaming. But AMD fared no better. Its proudest creation, the Athlon 64, was made up of 106 million transistors, still less than half the number used in ATI and NVIDIA’s DirectX 9 GPUs.

Transistor count, by itself, tells little of performance. A more useful measure is programmable floating point performance, or the speed at which a chip can do decimal arithmetic. Intel’s latest desktop CPU, the Core 2 Extreme X6800, can push an unprecedented 47 gigaflops a second. Next to ATI’s RADEON X1900 XTX, which can crunch 416 gigaflops, the CPU appears meagre and underwhelming.

To say that this is an astonishing development in computer architecture is an understatement; how did the ‘central’ processing unit become dwarfed by a mere add-on board?

Graphics processors have managed to soak up more transistors mainly due to their highly scalable architecture; because rendering performance increases linearly with the number of graphics pipelines, graphics processors used as many transistors as feasible to build multiple pipelines. CPU performance on the other hand does not generally scale linearly with the number of cores. As such, transistor resources have been devoted to cache and deepening the pipeline, which uses far fewer transistors than building additional pipelines. The ability of graphics processors to employ more pipelines is the reason why they have soaked up transistors faster than their CPU counterparts, which has in turn made them much larger.



With additional graphics pipelines come additional floating point units. So as GPUs scaled their pipelines to eight vertex units and 48 pixel units, CPUs stood still with their lone SIMD (Single Instruction, Multiple Data) unit, their performance only better than the CPU of yesteryear due to a minor bump in clock speed.

Various add-on boards have also boasted superior performance to CPUs at a specific task. What differentiates the GPU is that, over time, it has become increasingly programmable, to the extent that it can run a variety of programs, many unrelated to graphics, much faster than the CPU. This created the movement that is now known as GPGPU or General Purpose Computing on Graphics Processing Units.

Moving on up
The GPGPU movement did not come about overnight. Three key innovations made GPGPU possible. The release of high level shading languages, first with NVIDIA’s Cg, then Microsoft’s HLSL and OpenGL’s GL Shading Language made GPU programming accessible to the ordinary programmers. Vastly expanded GPU programmability, especially from Shader Model 3.0, brought support for loops, branches, large instruction counts and the 32-bit floating point data format. This allowed the GPU to mirror almost all of the functionality of the CPU. And finally is the adoption of PCI Express, which gave the GPU a high-bandwidth interface to write results back to main memory, something that AGP could never do.

What’s the purpose?
The purpose of this article was to survey the current landscape of GPGPU applications. But as we started to learn about all the research efforts under way, it became clear that there were not just a handful of GPGPU applications, but a whole smorgasbord. Physics acceleration, global illumination, protein folding, neural nets and SQL queries can all be accelerated on the GPU. Even something as remote as stock options pricing has been made to run on the graphics processor. To find out how the GPU became so versatile, let’s take a look at how GPGPU actually works.