Slashdot Mirror
Impressive GPU Numbers From Folding@Home
ludd1t3 writes, "The Folding@Home project has put forth some impressive performance numbers with the GPU client that's designed to work with the ATI X1900. According to the client statistics, there are 448 registered GPUs that produce 29 TFLOPS. Those 448 GPUs outperform the combined 25,050 CPUs registered by the Linux and Mac OS clients. Ouch! Are ASICs really that much better than general-purpose circuits? If so, does that mean that IBM was right all along with their AS/400, iSeries product which makes heavy use of ASICs?"
Yes. That's basically right.
Here's a Wikipedia article on general purpose GPU processing.
Folding is what's know as a rediculously parallel problem. That is, it can be broken up in to small subproblems that can be distributed among many processors with a minimal amount of communication among processors. It also benefits from not requiring a lot of branching (if/switch statements and such), which GPUs generally do not handle well.
Many problems, (I'd argue MOST problems) do not cater well to these kinds of restrictions. So, while a GPU is well suited to crunching away on pieces of the folding problem, it's going to be lousy at doing the day-to-day stuff you do with your computer.
It has nothing to do with memory bandwidth or use. The ASIC is about 1000 times faster than the CPU because it is using dedicated hardware designed to run very fast and parallel in 3D image processing, which is almost exactly the same problem as folding protiens.
Unless you are saying all CPUs are pegged at 99.9% use, or the GPU has memory three orders of magnitude faster then you're just looking at a effects that make a few percent difference here and there. The simple fact is the GPU is insanely faster at solving specific problems (3D processing) while it simply cannot ever run an operating system.
Fear: When you see B8 00 4C CD 21 and know what it means
X1900 - 48 pixel shader processors plus 8 vertex shaders. Assuming you manage to run them all equally in parallel: 56 processors.
Standard CPU - 1 core (assuming dual cores get read as 2 CPUs).
448 GPUs x 56 = 25,088 effective processors all with on card memory.
25,050 CPUs x 1 core = 25,050 effective processors all dealing with system busses etc.
In short, if you're performing one simple task trillions of times, many very simple, highly optimized processors with dedicated memory do the job better than even a similar number of much more capable processors that have to play nice across a whole system.
And this ignores the number of old couple of hundred megahertz systems that people don't use anymore so hand over to the task vs. X1900s being the very high end of ATIs most recent line.
For massively parallel tasks like rendering pixels, folding proteins, compressing frames of a movie, etc. I'd absolutely love large quantities of a simple processor. For most other tasks, given present technology, I'd still side with fewer more able processors. Either way comparing 448 of something with 56 processors within it to 25,000 single processors and saying, "But 448 is SO much less than 25,000!" is an unfair comparrison.
The technical term (jargon) is embarrassingly parallel.