The rise of the GPU

Posted on 10 Nov 2009 at 15:25

• 
• 
• 
• 
• 

 1 of 5

Gallery

Anatomy of a GPU

Although primarily intended for the very specific application of generating 3D graphics, a GPU is ultimately just a processor that carries out tasks that involve logic and maths. In this way, it's similar to a mainstream microprocessor. To understand what makes a GPU tick, we must look at the ways in which it differs from its better-known cousin, the general-purpose microprocessor. For brevity, we'll refer to this as the CPU (central processing unit).

The most obvious difference lies in the number of cores. The x86 CPUs that power our PCs currently have one, two, three, four or six cores, while Core i7 uses Hyper-Threading to actually double its cores. However, top industry engineers don't believe ordinary CPUs will go much beyond this in the short term. That's because, for general-purpose computing, the law of diminishing returns comes into play. This is due to the fact that it becomes more difficult to farm out work to all the cores in an effective way as their number increases.

To give an analogy, a production line at a pizza restaurant might work well with four people each handling a specific task: rolling out the dough, adding tomato, sprinkling cheese and adding other toppings. Adding a fifth member might actually slow the process down rather than speed it up, due to the extra management involved in delegating the various tasks and avoiding bottlenecks.

Parallel processing

Other jobs are parallel in nature. For example, in a pea-shelling factory, adding extra workers gives a productivity boost because the management involved in scaling up the workforce is minimal. 3D graphics processing falls into this category, and that's why it's possible to employ multiple cores effectively.

GPUs, therefore, have tens or even hundreds of cores, with the top spot currently going to the AMD GPU that's used on some of the ATI Radeon-branded graphics cards, weighing in at 800 cores. However, since top-of-the-range GPUs have a similar number of transistors to top-of-the-range CPUs, it's clear that the cores in a GPU must be radically different from those in a CPU.

The major differences lie in the type of code the two types of processor are designed to execute. CPUs are primarily aimed at office-type applications where If-Then-Else logic statements are common. In these circumstances, the code often doesn't execute in strict sequence. Instead, it branches out along one of many possible permutations. Programmers refer to this as branching code. General-purpose software also tends to include loops whereby it executes the same code over and over again, and reuses small amounts of data.

Both these factors result in transistor-hungry circuitry to improve the performance of general-purpose software, as Nick Stam, Nvidia's technical marketing director, explains. "CPUs include complex branch-prediction hardware to process branching code effectively. They also have large onboard caches that can store the code loops and small datasets, allowing high performance compared to having to access such code or data from system RAM."

Multiple-core GPus

While branches also occur in GPU code, they are less frequent so they don't need a large amount of branch-prediction hardware. This means that a GPU can use more of its transistors implementing the hardware to actually perform the processing. So rather than having a few complex cores, a GPU has simpler cores, but many more of them. According to Nick Stam, "This delivers high performance for media-rich data types, including 3D graphics data, because this type of media stream typically uses independent data elements, such as screen pixels, that can be processed in parallel with every clock tick of the GPU".