FordGT90ConceptI wonder how much he has to do with NVIDIA being in cahoots with the Folding @ Home project.

FordGT90ConceptWhat this guy is liable to do is remove NVIDIA from the gaming market altogether by striving to increase folding performance. It's already happening too seeing how many people build computers with 2+ NVIDIA cards in it just for folding. I really don't like where NVIDIA is going with this, hence my comment about NVIDIA potentially losing a customer.

I'm just glad Intel is getting ready to enter the market with NVIDIA perhaps leaving.

FordGT90ConceptWhat they are doing is not philanthropic. What they're doing is capitalizing on philanthropy. If NVIDIA was actually being philanthropic here, they would design a card specifically for folding and create a large farm just to donate to the project. They aren't doing that.

They play the middle man. I got these cards which are supposed to be great for gaming but you can also use them to simulate protein folding for Stanford. The more you buy and the more you run them, the higher your score. What, exactly, is NVIDIA doing philanthropic except facilitating the movement of more product?

FordGT90ConceptI've been down this road before and it's practically arguing religion ("but it cures cancer!!!!"). There's no sense in continuing.

Cancer is nature's way of saying you've out lived your usefulness.

FordGT90ConceptWhat they are doing is not philanthropic. What they're doing is capitalizing on philanthropy. If NVIDIA was actually being philanthropic here, they would design a card specifically for folding and create a large farm just to donate to the project. They aren't doing that.

They play the middle man. I got these cards which are supposed to be great for gaming but you can also use them to simulate protein folding for Stanford. The more you buy and the more you run them, the higher your score. What, exactly, is NVIDIA doing philanthropic except facilitating the movement of more product?

Why not just use a supercomputer?

Modern supercomputers are essentially clusters of hundreds of processors linked by fast networking. The speed of these processors is comparable to (and often slower than) those found in PCs! Thus, if an algorithm (like ours) does not need the fast networking, it will run just as fast on a supercluster as a supercomputer. However, our application needs not the hundreds of processors found in modern supercomputers, but hundreds of thousands of processors. Hence, the calculations performed on Folding@home would not be possible by any other means! Moreover, even if we were given exclusive access to all of the supercomputers in the world, we would still have fewer computing cycles than we do with the Folding@home cluster! This is possible since PC processors are now very fast and there are hundreds of millions of PCs sitting idle in the world.

FordGT90ConceptCancer is nature's way of saying you've out lived your welcome.

DarkMatterI think you don't understand what F@H is. No company can build a fast enough supercomputer, Nvidia by pushing GPGPU and F@H, and by teaching GPGPU in universities is doing much more than what a farm of supercomputers can do.

DarkMatterJust for an easy comparison. Fastest supercomputer Roadrunner has 12,960 IBM PowerXCell 8i CPUs and 6,480 AMD Opteron dual-core processors and a peak of 1.7 petaflops. Now looking at the statistics in these forums I find there are 38,933 members. If only half the members contributed to F@H at the same time, there would be much more power there. Now extrapolate to the world...

DaedalusHeliosA powerful GPU is also a powerful folding@home card. A weak Folding@home card is also a weak GPU... the properties that make a good GPU also make it good for folding if that makes since. Thinking they are seperate things and it might comprimise graphics performance is a non-issue so don't think it will cause a problem.

FordGT90ConceptGPGPU is fundamentally wrong. Intel's approach is correct in that there's no reason GPUs can't handle x86 instructions. So, don't teach proprietary GPGPU code for NVIDIA's profit in school, teach them how to make GPUs effective at meeting D3D and x86 requirements.

Those computers have extremely highspeed interconnects which allows them to reach those phenominal numbers; moreover, they aren't overclocked and they are monitored 24/7 for problems making them highly reliable. Lots of people here have their computers overclocked which breeds incorrect results. If that was not enough, GPUs are far more likely to produce bad results than CPUs.

There obviously are inherint problems with Internet-based supercomputing and there's also a whole lot of x-factors that ruin it's potential for science (especially machine stability). Folding especially is very vulnerable to error because every set of work completed is expanded by another and another. For instance, how do we know that the exit tunnel is not the result of an uncaught computational error early on?

FordGT90ConceptAs was just stated, a 4850 is just as good as the 9800 GTX in terms of gaming but because of the 9800 GTX's architecture, the 9800 GTX is much faster at folding. This is mostly because NVIDIA uses far more transistors which means higher power consumption while AMD takes a smarter-is-better approach using far fewer transistors.

And yes, prioritizing on GPUs leaves much to be desired. I think I recall trying to play Mass Effect while the GPU client was folding and it was unplayable. It is a major issue for everyone that buys cards to game.

DarkMatterThe same aplies to x86. The difficulty relies on making the code highly parallel. x86 is NOT designed for parallelism and is as difficult making a highly parallel computing program in x86 as doing it in GPGPU codes.

DarkMatterFalse. GPGPU is as prone to errors as supercomputers are, they doublecheck the data is correct in the algorithms. Even if that takes more computing time, reducing efficiency, beause the seer computing power of F@H is like 1000 times that of a supercomputer that means squat.

A GPU does not make more errors than a CPU anyway. And errors resulted from OC yield highly unexpected results that are easy to detect.

Anyway F@H is SCIENCE, do you honestly believe they only send each algorithm to a single person?? The have 1000's of them and they know which one is well and which not. :laugh:

DarkMatterG92 (9800 GTX) has much less transistors than RV770 (HD4850) FYI. And the 55nm G92b variant is significantly enough smaller too 230mm^2 vs 260mm^2.

DarkMatterOf course folding at the same time reduces performance, but the fact that GPGPU exists doesn' make the card slower. :laugh:

FordGT90ConceptIntel is addressing that.



F@H doesn't double check results.

What happens when a CPU errors? BSOD
What happens when a GPU errors? Artifact

Which is fatal, which isn't? CPUs by design are meant to be precision instruments. One little failure and all goes to waste. GPUs though, they can work with multiple minor failures.

I got no indication from them that any given peice of work is completed more than once for the sake of validation.


No, errors aren't always easy to catch.
Float 2: 00000000000000000000000001000000
Float 4: 00000000000000001000000001000000

If the 17th digit got stuck, every subsequent calculation will be off by a minute amount. For instance:
Should be 2.000061: 00000000000000010000000001000000
Got: 4.0001221: 00000000000000011000000001000000

Considering F@H relies on a lot of multiplication, that alone could create your "exit tunnel."

9800 GTX = 754 million transistors
4850 = 666 million transistors

Process doesn't matter except in physical dimensions. The transistor count only changes with architectural changes.

It's poorly executed and as a result, CUDA is not for gamers in the slightest.

DarkMatterIt's SICIENCE so of course they have multiple instances of the same problem. They don't have to say that because they are firstmost and ultimately scientists working for scientists.

DarkMatterWRONG. RV670 has 666 m transistors. RV770 has 956 m transistors. source
source

DarkMatterOf course it's not for games (except for PhysX). But it doesn't interfere at all with games performance. Intel's Larrabee, x86 or not, won't help at games neither. In fact there's no worse example than Larrabee, for what you are trying to say. There'd not be a GPU worse at gaming than Larrabee.

FordGT90ConceptPande is a chemical biologist. How much he cares about computational accuracy remains to be seen.

So, Tom's Hardware is wrong. And? That doesn't change the fact that F@H prefers NVIDIA's architecture.

That makes a whole lot of no sense so I'll respond to what I think you're saying.

-NVIDIA GeForce is designed specifically for Direct3D (or was).
-CUDA was intended to offload any high FLOP transaction from the CPU. It doesn't matter what the work actually is comprised of.
-CUDA interferes enormously with game performance because it's horrible at prioritizing threads.
-Larrabee is a graphics card--but not really. It is simply designed to be a high FLOP, general purpose card that can be used for graphics among other things. Larrabee is an x86 approach to the high-FLOP needs (programmable cores).

Let's just say CUDA is riddled with a lot of problems that Larrabee is very likely to address. CUDA is a short-term answer to a long term problem.

HaytchIf CUDA doesnt increase f.p.s, nor does it decrease it. Then that equals even.

FordGT90ConceptIf something is using CUDA while a game is running, it hurts the game's FPS badly.

DarkMatterI can see your disbelief about science, but I don't condone it. Scientists know how to make their work, assuming they don't is plainly stupid.

DarkMatterMany people think that GPGPU is the BEST answer for that, and they all of them don't work for Nvidia. In fact, many work for Ati.

DarkMatterBut NOTHING forces you to use CUDA at the same time, that's the point. When you are gaimng dissable F@H, of course!! But when you are not using the GPU for anything you can fold, and with GPU2 and Nvidia card you can fold MORE. It's simple.