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?"

Are ASICs really that much better?

[by Anonymous Coward]

Are ASICs really that much better than general-purpose circuits?

Generally ASICs are much better than general-purpose circuits except in general cases.

Re: Are ASICs really that much better?

[Jeffrey Baker (6191)]

Good one ... but I also wonder why anyone is throwing around the term "ASIC" in this article. A GPU is obviously not an application-specific circuit, which is clearly shown by the fact that it can be programmed to process graphics, or protein folding, or numerous other tasks. A GPU is a general-purpose processor like a CPU, it just happens to have different numbers and kinds of execution units.

Re: Are ASICs really that much better?

[doublebackslash (702979)]

Partially true. The GPUs of today now have some general purpose circuits, but they are far from optimized and the execution unit count is skewed to the point that these processors would never, ever be able to run, say, an OS with anything approaching efficiency. FAH benefits from the insane amount of Floating point power because FAH is nothing but a pure FP stress test. They had to heavily modify the code to run on these babies, basically tuning the problems into vector information and letting the GPU do its thing, throwing. Only a few areas involve a need for CPU style processor, which is functionality provided only on these new cards. So please, please realize that even though these cards do not a contain a "protein folding circuit", they did modify the program to run on what it does have: 4x4 matrix operation units for multiplication and addiction.

IBM right all along, or obvious?

[Gothmolly (148874)]

Custom app written to run on hardware specifically designed to run apps like it, outperforming general purpose CPUs? Newsflash from Ric Romero!!1!

Distributed amongst home users

[Skevin (16048)]

So, will someone please create a really pretty 3D screensaver representing the folding calculation process? I'd love to see a represention with hi-res lighting and texturing, full transforms, and user-scalable views at 400 million triangles/sec.. Thanks.

Solomon

Re:Distributed amongst home users

[Enderandrew (866215)]

The folding team has done this, and it will be a free download for the PS3 version. The Cell processor runs the Folding application itself, and the graphical representation of the protein folding calculations will be handled by the GPU with a pretty display.

GPUs are Specialized Parallel Computers

[ThinkFr33ly (902481)]

GPUs are, for the most part, highly specialized parallel computers [wikipedia.org]. Virtually all modern CPUs are serial computers. They do essentially one thing at a time. Because of this, most modern programming languages are taylored to this serial processing.

Making a general purpose parallel computer is very, very hard. It just so happens that you can use things like shaders for more than just graphics processing, and so via OpenGL and DirectX you can make GPUs do some nifty things.

In theory, and indeed often in practice, parallel computers are much, much faster than their serial counterparts. Hence the reason a GPU that costs $200 can render incredible 3D scenes that a $1000 CPU wouldn't have a prayer trying to render.

Re:GPUs are Specialized Parallel Computers

[dslauson (914147)]

Yes. That's basically right.

Here's a Wikipedia article [wikipedia.org] 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.

Re:GPUs are Specialized Parallel Computers

[Goner (5704)]

The technical term (jargon) is embarrassingly parallel [wikipedia.org].

This is the perfect time...

[loraksus (171574)]

... to start heating your house with your computers ;)

I actually installed boinc with seti on several of my machines last night and it worked quite well to heat part of the house (us Canadians need to turn the heater on earlier). Took a bit of time to get started, but it was nice and toasty in the morning.

Does anyone know if this method is less efficient in generating heat than using a apace heater? Slower perhaps...
If you're going to use energy by turning on the wall heater anyways, why not use it to crunch some numbers?

Re:This is the perfect time...

[Murphy Murph (833008)]

I actually installed boinc with seti on several of my machines last night and it worked quite well to heat part of the house (us Canadians need to turn the heater on earlier). Took a bit of time to get started, but it was nice and toasty in the morning. Does anyone know if this method is less efficient in generating heat than using a apace heater? Slower perhaps..

Using your CPU as a space heater is not a bad idea. It is 100% efficient. Every watt it consumes gets turned into heat. Before someone says "but the cooling fans are wasteful" let me remind you that the air moved by those cooling fans will eventually come to a stop (inside your house) as a result of friction, releasing its energy as heat in the process.

Depending on what type of space heater you use, and the construction of your house, your computer can be more efficient than many other electric space heaters. Since none of the energy "consumed" by your CPU/GPU is converted to visible light, none of it has the opportunity to leave your house through your window panes (assuming you have IR reflective glass). Contrast this to quartz and halogen space heaters which produce a fair amount of visible light.

In much the same way, incandescent bulbs match the efficiency of compact fluorescents during the winter months. Every watt "wasted" as heat during the summer is now performing useful work heating your house. (Before someone says "you called a quartz/halogen space heater inefficient because of its waste light, and now an incandescent efficient because of its waste heat!' let me say that the space heater's light is not useful light, while the bulb's heat is useful heat (during the cool months.))

Re:This is the perfect time...

[evilviper (135110)]

It is 100% efficient.

Not true. You aren't taking into account Power Factor at all... Not that I'm surprised, as most people don't understand it.

With switching power supplies, it's common to see PF in the range of 0.4, as opposed to fully-resistive electric space-heaters (and incandesent lightbulbs) with a perfect 1.0 PF.

Residential customers are lucky, in that they don't get charged for PF losses by the power company, while companies certainly do. However, it's still highly ineffecient, even if you aren't paying for it directly.

And besides that, electric heating is almost always more expensive than conventional heating, like natural gas, or electric heatpumps.

Not really.

[TerranFury (726743)]

>Using your CPU as a space heater is not a bad idea. It is 100% efficient.

Not really. Consider exergy [wikipedia.org]. Yes, your CPU is just as efficient as any electric space heater. However, consider that the alternative is probably burning natural gas or oil in a furnace. If you burn fuel for heat, 90%+ of the chemical energy goes to producing heat (the rest is lost as unburnt hydrocarbons in the exhaust). If you burn fuel to spin a turbine at a power plant, only about 40% goes to electrical energy, and unless it's a cogeneration plant which uses the waste heat for industrial purposes, the rest is lost as heat up the smokestacks. So, starting from the fossil-fuel source, electrical heating is less than half as efficient as burning fuel for heat. If you do need to heat using electric power, it's much more efficient to use that electricity to pump heat in from a lower temperature outside than it is to turn that electricity itself into heat.

If you are stuck with electric (non-heat-pump) heating in your house, however, you are correct: There is absolutely no reason not to run your CPU or any other electrical appliance full tilt.

Re:This is the perfect time...

[ameline (771895)]

Ok, I went and did the math (assuming, on average 1035 btu/cubic foot of natural gas) Looking at my bills,

Natural gas is (cdn)$0.278287 per cubic meter, and electricity is 0.058 /kwh.

At 96% efficiency, natural gas works out to 0.027331 / kwh, (3413 btu in 1 kwh) or 47% of the cost of electricity at today's prices in Toronto.

So 1/3 was a bit of hyperbole, but not too much.

Re:

[loraksus (171574)]

Well, I really, really hate the "heater that hasn't been used for 6 months" smell, so that was sort of my primary focus.
That makes it a little better... right?
Please?

Move the vector processor on-board?

[Zygfryd (856098)]

So when are we going to see (x86/64) motherboards with a socket for a standard processor and a socket for a vector processor?
Couldn't we finally have graphics cards that only give output to the screen and separate vector processors with a standardized interface / instruction set?

Remember: 1 GPU has more than one processor.

[nick_davison (217681)]

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.

Actually Linux users have good hardware:

[olddoc (152678)]

Look at the number of Tflops per active cpu by OS.
I took (TFlop/active cpu)*1000 to get a readable number --
or Gflops/cpu
Windows is .948
Mac is .51
Linux is 1.21
And GPU is 65!

The source:
http://fah-web.stanford.edu/cgi-bin/main.py?qtype= osstats [stanford.edu]

The average Linux user proably has a decent AMD Athlon,
The average Windoze user has a P4 Dell.
Athlons just crunch the math better.

Re:So obvious...

[Iron Condor (964856)]

ASCII silly question, get a silly ANSI.

Re:So obvious...

[Jesus_666 (702802)]

UTF are you talking about? I'm quite sure the mods are not latin-1 post like this go unmoderated.

Re:Lopsided Alright..

[blahplusplus (757119)]

I have a feeling this is memory bandwidth related, modern GPU's have insane amounts of memory bandwidth compared to the wide range of desktops. Not to mention the parallelism.

Re:Lopsided Alright..

[Majik Sheff (930627)]

Look at the first two letters of the acronym: Application Specific. A screwdriver and a swiss army knife will both turn a screw, but the screwdriver is going to be much more efficient at it. GPUs are finely tuned to rip through massive volumes of floating point vectors and not much else. It just so happens that the folding project also fits this desctiption and as such is an excellent use of an otherwise wasted resource.

Re:Lopsided Alright..

[throx (42621)]

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.

Re:Lopsided Alright..

[oringo (848629)]

You can look at the statistics many ways. Here's the GFLOP/CPU catagorized by OS:

1. GPU: 65.463
2. Linux: 1.219
3. Windows: 0.948
4. Mac: 0.511

Of course, GPU beating the hell out of CPU in such tests is no surprise. It's pretty much a massive parallel vector engine. I'm more interested in seeing how PS3 holds agains all other guys when it comes out. They have a folding client for PS3 already.

Re:Not a mystery

[tkittel (619119)]

Your logic is fine, but you are overestimating the effect you mention if you really think that it "solves the mystery".

500 users out of 25000 means that you have at most taken the 2 percent highest performers out of the CPU pool. If we assume that those 2 percent have computers that are 5 times as powerful as the average computer, then we have lowered the average performance of the CPU pool by roughly 9%.

This 9% systematic effect will lower the reported performance superiority of around 5000% of the GPU vs. the CPU to something like 4500%. I.e. it doesnt change the result at all (which seems to be that GPUs kick ass for these applications).