ATI and NVIDIA stats vs. PPD numbers

[P5-133XL]

I'm actually not suggesting that the Point system be changed at all. What I'm saying is that the stats indicate that the amount of calculating done by the NVidia and ATI GPU2 clients are very close while the PPD is very seperated (around 2.5X) and that does not mesh with what should be happening. People here don't like the concept that the PPD calculations are off (they threaten to lock the topic, if the thread goes down that path), so I'm now proposing that the Stat pages must be wrong as the other alternative. If the client stat page is wrong, then what can be done to correct it and if it can not be corrected then why does it exist because wrong data is far worse than no data.

Personally, I don't know which is wrong PPD, Stat, or me. However, it is obvious that the PPD calculations can't be wrong because people here are not going to tolerate that concept. I'm still waiting for some argument that shows me that I'm wrong and that TFLOPS and PPD for the two GPU2 clients actually should not correlate but rather diverge. So, at this point, it must be that the TFLOPS on the stat pages are being miscalculated.

[MtM]

Thread should have been clossed

[P5-133XL]

Thread should have been clossed

It can still be done, all it takes is a whim by one of the many moderators.

[MtM]

Yes indeed.

Edit: on the other hand, if someone does have a worked out idea this is the time and place to post it

[KezNews]

"*TFLOPS is the actual teraflops from the software cores, not the peak values from CPU/GPU/PS3 specs."

So it means effective TFLOP done in GPU core = amt of work done.

Me think people forget ATI people most 3850,3870(x2),4850,4870(x2), so PPD per card high, while Nvidia client work on many low end one like 8500GT and 8600GT also. So super-high PPD GTX280 balanced by lower end 8 series.

[Ren02]

I vaguely remember that GPU1 client used to calculate the forces between atoms twice. No idea if that has been fixed in the GPU2 client.

During a single simulation step the first atom is picked and the forces between this atom and all the other atoms are calculated and summed. The resulting vector shows where this atom will move for the next step.
The process is repeated for the second atom. At this point the force between first atom and second atom has in fact already been calculated but it is difficult to retrieve it from the memory and therefore it is faster to recalculate the force between second atom and first atom again. When we move to atom three then forces between first and third and second and third have already been calculated. When we get to the last atom, all the forces between it and previous atoms have been calculated. On the average every second floating point operations is therefore "wasted".

Could it be that Nvidia has found a way to retrieve the previous calculations from the memory and therefore do 2 times less work to simulate a step? If that's the case then the similar TFLOPS of Ati and Nvidia clients just show that Nvidia uses a more efficient algorithm.

[7im]

Look, AMD and Intel CPUs have ALWAYS performed differently, and even gotten different types of work units. What's the big deal that ATI and NV video cards perform differently also? CPU and GPU architecture very greatly, and so does the software that each uses to communicate with the fah client.

TFLOPS is a measure of calculations performed, NOT scientific contribution. TFLOPS is not a measure of client efficiency either, and never will be.

If you want to talk about points, again, I suppose we can do that. But leave TFLOPS out. They are nice to look at, but don't mean a much unless you have the decoder ring. And even then, it wouldn't translate well in to scientific relavance or contribution.

[P5-133XL]

The whole point of this discussion is TFLOPS vs PPD on the GPU2 Client. Earlier, PPD was forbiden as a subject. Now PPD is OK, but TFLOPS aren't. Is your goal just to interfere with the topic, so it can't be discussed intelligently?

If TFLOPS don't mean much, then whats the point of having a web page dedicated to it?

Just because you don't have the decoder ring, does not mean others shouldn't discuss the topic. Or, perhaps it does, because it is unacceptable that someone talk about something that the censor can't understand. I didn't realize that we were in such a closed society here.

P.S. 7im, I'm starting to really get annoyed with your attempts at censorship.

[MtM]

The whole point of this discussion is TFLOPS vs PPD on the GPU2 Client. Earlier, PPD was forbiden as a subject. Now PPD is OK, but TFLOPS aren't. Is your goal just to interfere with the topic, so it can't be discussed intelligently?

If TFLOPS don't mean much, then whats the point of having a web page dedicated to it?

Just because you don't have the decoder ring, does not mean others shouldn't discuss the topic. Or, perhaps it does, because it is unacceptable that someone talk about something that the censor can't understand. I didn't realize that we were in such a closed society here.

P.S. 7im, I'm starting to really get annoyed with your attempts at censorship.

He's not trying to censor, he's trying to avoid a repeating circle of old discussions, wrong facts ect. Instead of complaining why he does what he does, in your pov that is, why don't you ask him why he says what he says? You're bringing up points and I find the discussion going civil enough to assume you will get a civil and intelligent answer back. I know I did when I thought I had legit questions and remarks, this is the opposite of a closed society but it's the attitude which you carry with yourself which will close you off from people. I talk from experience, trust me

[Sahkuhnder]

Perhaps someone in the know could add to the WIKI information about both TFLOPS and PPD and their relationship to each other. Then in the future when the topic comes up again (and it will) the curious party can be referred to a source of knowledge.

[VijayPande]

If TFLOPS don't mean much, then whats the point of having a web page dedicated to it?

because it's the main metric used in supercomputing. It's a horrible metric there too (since Blue Gene flops != FAH flops -- BG would do much worse at FAH than their reported FLOPS), but it's a single number which makes it easy to compare.

Two codes can use very different flops to do the same science. The GPU code uses a lot of FLOPS for example to do the same science the CPU code would (often 2x more), since that's the fastest (in terms of wall clock) way to get the calculation done. Wall clock time is in the end the only thing that matters and that's why we benchmark on wall clock (and why in our papers, we emphasize wall clock).

Finally, you have asked why FLOPS and PPD diverge for the two GPU's. The problem with optimizing the ATI GPU is that it has a lot of FLOPS, but getting the bytes to the FPU's is the bottleneck. It doesn't matter if you *could* do a lot of FLOPS if you can't get the data there to do the calculation. In all computing these days, speed is a combination of FLOP power and memory access. In GPU's, memory access is the real issue these days, since all high end GPUs can do lots of FLOPS in an ideal setting. Here's an analogy for you: it doesn't matter if your top speed is 200MPH if you're on a road with lots of traffic lights where you constantly have to stop. In that case, even a slower car on a non-traffic light highway will do much better.

[DreadedOne509]

Had to giggle to myself about that analogy, thanks for the smile!

Considering the above statement by Dr. Pande, and assuming I understood even half of what he just said, would a GPU card with faster, and more memory
generally perform better given that all else is equal?

[7im]

I think you might have misunderstood the memory issue slightly. It's the getting the bytes in to and out of the card, not necessarily storing the bytes. So as many have noted in the PPD threads (which are still open and running P5-133XL) overclocking memory doesn't increase performance much, if at all. But increasing the shader and core speeds do increase performance considerably.

For reference, if I curtail a dicussion, it's usually because it's either a dead horse, or a flame war waiting to ignite. PPD comparison threads are typically both.

[codysluder]

Had to giggle to myself about that analogy, thanks for the smile!

Considering the above statement by Dr. Pande, and assuming I understood even half of what he just said, would a GPU card with faster, and more memory generally perform better given that all else is equal?

Maybe, but I doubt it.

I'm certainly not an expert on these matters, but getting the data to the processor is more than just on-board vram speed. The data has to be moved between main ram and vram (pcie speeds), a program segment must be issued to each parallel processor with an appropriate data pointer, the FLOPs must be executed producing the new values stored in a new segment of vram, the step counter must be updated as the new data gets moved to the old data array, and the process begins again. (It's not clear what data must move through the pcie buss and when.)

Depending on how many program segments are generated compared to the number of parallel processors, additional time may be spent restarting additional program segments within the outer "step" loop. Allocation of general registers from a global pool may also complicate this process.

These steps may be similar between different GPU architectures, but they cannot be identical ... which is the whole point of different architectures. In the long run, FAH will run differently on ATI an Nvidia and small proteins will run differently than large ones, perhaps even increasing the throughput on one while decreasing the throughput on the other.

There's a lot more going on than just vram size/speed, though those both probably have some influence. Vijay's traffic lights matter, too. It's just that we don't know what constitutes a GPU traffic light.

[scott.ager]

If TFLOPS don't mean much, then whats the point of having a web page dedicated to it?

because it's the main metric used in supercomputing. It's a horrible metric there too (since Blue Gene flops != FAH flops -- BG would do much worse at FAH than their reported FLOPS), but it's a single number which makes it easy to compare...

C'mon !, Go ahead and tell them. The web page is a distraction to occupy people that don't know how to do biochemistry