Using GPUs for general purpose processing

[Asher]

http://gamma.cs.unc.edu/GPUSORT/results.html

A bunch of beanies at UNC's CS department have figured out how to use a GPU (Graphics Processing Unit) to do some sorting, and lo and behold, it's much faster than a general purpose CPU.







It'll be pretty impressive, I'd imagine next-gen consoles will be able to use these first. Xenos on the Xbox 360 is even more generalized than any current other GPU, so it can possibly do even more. It essentially has 48 ALUs and Vector processors on the thing.

[Tattila the Hun]

Faster turns in CIV 4?

[DanS]

This is cool. I am involved in a project that looked at this several years ago and determined that GPUs did a lot of rounding, which made them unsuitable for calculations needing precision. But maybe we were wrong.

[Asher]

New architectures can do 32-bit float vectors easily.

[DanS]

If this is so, then a GPU could be an excellent complement to an Intel chip.

[Asher]

The new shader instruction languages also include support for looping, branching, etc. New GPUs also have full access to the system RAM and CPU via PCI-Express (16x, so 4GB/s up and 4GB/s down simultaneously).

[Verto]

Asher is as constant as the north star in creating the most boring thread discussions on Apolyton.

[Asher]

It's only boring if you don't understand the implications.

The trend is to distributing workload away from the CPU.

There is now a Physics Processing Unit for PCI-Express also.

[Verto]

OMG ORGASM!!!

[JimmyCracksCorn]

I guess the question I have is when can we expect the GPUs we have to be utilized in things like making Civ turns faster? Is it a software thing, or does hardware need to be changed in order for it to start being used?

[Asher]

It's the software thing. Since Civ4 is a 3D game, I doubt they'd use much (if any) of the GPU for that kind of thing. They probably want the resources for graphics.

[Asher]

http://www.tomshardware.com/hardnews/20050630_161353.html

Graphics processors supercharge everyday apps

By Scott Fulton

June 30, 2005 - 16:13 EST

Chapel Hill (NC) - Originally developed to remove a massive processing workload from the CPU, some scientists examine how the graphics processors can accelerate non-graphic applications as well. The Geometric Algorithms for Modeling, Motion, and Animation (GAMMA) Research Group at the University of North Carolina at Chapel Hill, reported this week that Nvidia's 7800 GTX reference card increased the speed of test applications by up to 35x.

Researchers with GAMMA said they found enormous performance capabilities in Nvidia's newest graphics card that substantially outpaces its predecessor, the GeForce 6800 Ultra. Compared to the 6800, the 7800 tripled performance; without the help of a graphics chip, the speed gains were between 8x and up to 35x. The discovery points to the removal of a bandwidth bottleneck, which may lead to the unimpeded development of co-processing libraries and software development kits (SDKs) for everyday applications, such as spreadsheets and database management systems.

Click here for Maxtor

"It seems to me that the floating-point bandwidth on the new hardware is much more than on a 6800 Ultra," reported Naga K. Govindaraju, Research Assistant Professor at UNC's Department of Computer Science, in an interview with Tom's Hardware Guide. "On a 6800 Ultra, we are, in some manners, very limited...Since the bandwidth is not good enough on the card, we were still not able to use the full performance of the card. On a 7800 GTX, it seems to me that the floating-point bandwidth is much higher."

graph The trick to exploiting the latent power of the graphics processor while it isn't producing scenery for 3D games, UNC Professor Dinesh Manocha told us, is to rephrase everyday operations as though they were specific two-dimensional graphics functions, like texture mapping. While everyday CPUs work with threads, Prof. Minocha pointed out, graphics processors deal with streams capable of performing single instructions on multiple data elements simultaneously, through pipelines. By comparison, CPU-based parallelism divides instruction threads among multiple cores, for what Prof. Minocha calls a "von Neumann bottleneck." The NVIDIA 7800 GTX utilizes 24 pixel pipelines and eight vertex units for its implementation of Single Instruction / Multiple Data (SIMD) architecture.

This technique of essentially pretending everything is a game, stated Prof. Govindaraju, reduces the critical elements of such everyday functions as sorting algorithms to a single instruction, which the graphics processor then applies to multiple pipelines at once. Recent test results presented by the GAMMA team compared the performance of their GPUsort algorithm to a traditional linear QuickSort algorithm, compiled first under Microsoft Visual C++, then under Intel's C++, which is optimized for Hyperthreading. For sorting an array of 18 million elements, the Visual C++ routine required about 21 seconds to accomplish what the GPUsort routine produced in under 2. Hyperthreading and the Intel compiler boosted QuickSort performance to about 17 seconds.

One of the purposes of the GAMMA team's work is to demonstrate the extent to which computing power in everyday PCs lies dormant, especially with regard to mere productivity applications as opposed to computation-rich 3D games. Profs. Manocha and Govindaraju agree that general purpose computation libraries for such programs as Excel and Matlab could be the first step to the future development of SDKs that make full-time use of graphics chips as math coprocessors.

But what Prof. Manocha also pointed out is that the performance increase in GPUs is exceeding the rate of CPUs. "If you look at [both] computation power and rasterization power," stated Prof. Manocha, "in the last six years, [performance for] PC graphics cards has grown at a [factor] of 2 or 2.25 per year, whereas CPUs are barely doubling every 18 months." He added that he expects this trend to continue as both ATI and NVIDIA produce their next generations of graphics cards in 2006.