Wu 17721 on Radeon 6900XT (WTH???)

[Starman157]

My perception of the WU set is based on the PPD. "Normally" the 7950 generates about 260,000 PPD. With the problematic WU set mentioned here, that number is around 110,000, hence my 2.5x approximate slowdown.

[Starman157]

Ok, Joe. I get it now. FAH categorizes the GPU's based on incomplete scanning of the PCI device ID, and based on that categorization, you are correct, turning off 6900XT processing of this problem set will turn off all NAVI based work.

Thanks for the clarification.

[bruce]

FAH categorizes the GPU's based on incomplete scanning of the PCI device ID, and based on that categorization, you are correct, turning off 6900XT processing of this problem set will turn off all NAVI based work.

AMD didn't make that any easier. When GPUSpecies was adopted as the criterial it was easy to differentiate the manufacturer's code and the assigned device code. AMD categorized their devices by those two number plus a third one that FAH ignores.

All of "Navi 21 [Radeon RX 6800/6800 XT / 6900 XT" have the same classification based on the device code 0x73bf and FAH can't tell them apart. They all have the same device code.

I'll see if we can get P17721 blocked from AMD. It's also possible that the PPD for the nV GPU is too high.

As I said earlier, FAH is going to be redesign the species concept, but that project seems to have gotten stalled.

[bruce]

In fact, all nV GPUs get an "unfair" advantage over AMD. NV GPUs can run CUDA code which has been highly optimized by nV. AMD has not licensed CUDA (and probably won't, since it's owned by nV and would be expensive). AMD is limited to OpenCL, which runs slower.

FAH supports whatever either one supports.

Are both in 16x slots?

The 3080 is rated at 29.77 FP32 TFLOPS and the 6900XT is rated at FP32 23.04 TFLOPS
(In fact, FP64 performance is quite different, too, favoring nV by a large factor but I don't use that number since it makes a much smaller contribution to overall performance than FP32 does.)

Then, too, FAH awards a significant bonus for early returns so it's really an unfair comparison for several cumulative reasons.

[Starman157]

Interesting comments Bruce. Yes, I understand that the 0x73bf code pertains to the 6xxx series from AMD. Joe mentioned that the way the categorization process works involves lumping the 5xxx series into the same "bucket" even though the 0x731f code differs so turning off processing this set also includes the 5xxx series from AMD.

Blocking the P17721 is just one member of this set. I believe the full set ranges from 17711 to 17743 (the listing on the FAH project description lists 17711-17724).

As for the differences in the PPD awards, I'll leave it up to the FAH folks who make those decisions. Balance would be nice though.

Yes, I'm aware of the "unfair" advantage. Along with it comes the differences in price, availability and energy use.

As for 16x slots, mine sure is. PCI Express 4, now with resizable BAR. As for the bonus, yes, it would seem to be a cumulative issue.

[bruce]

We don't have any real information about the adjusted BAR with respect to FAH. Personally, I doubt it will matter much if it ever happens to be supported.

FAH doesn't create video so frame-rate means nothing.

[muziqaz]

5.5m from 6900xt is a pipe dream, unless you get some half a million atom sized project and you are not folding on your CPU. Other than that 6900xt averages from 2.8 to 3.5m PPD.
1.6m PPD for p17721 is too low. We did not see such PPD in the initial testing (and it was tested on 6900xt). I cannot check how much PPD exactly 6900xt get in testing, but if I don't forget I'll check when I get back home from work

P.S. Resizable BAR is disabled on test system

[Starman157]

I guess the pipe dream has arrived. I've seen 5.6m from my 6900xt at times while simultaneously folding on my 3950x. I believe that the atom count was around 270,000 or so, so well under your stated half a million. I'm presently seeing 5.2m reported (fluctuates from 4.7 to 5.2) from 17340 (prcg 17,412,22) with 257,383 atoms as I type this reply (the 3950x is simultaneously doing 589,789 on WU 16815 (prcg 16,686,23)).

I took a look at the highest rated WU for it's average. WU 14651 with 438,651 atoms in it; the RX6000 series average is 5,949,052. No mention the makeup of the actual cards used or whether there was simultaneous folding performed on the CPU (and what it was).

But that's the problem with averages. Averages that lump the 6800 (6700xt?) up to the 6900xt together as there is no discrimination as to the actual video card used. Is that average over all WUs and if so, what averages would you get with a diet of 5.0m and 1.6m units? What averages would you get with 4.0m and 1.0m units (or worse?) in the same "basket" courtesy of the slower cards in the same category?

As for the resizable BAR? I didn't expect it to have any effect on FAH (one could hope). I'm just stating that my BIOS is current, which includes current video drivers (21.3.2).

[muziqaz]

If the project produces 1.6m PPD on 6900xt, it gets excluded from AMD cards altogether, as chances are that project will run even worse on GCN architecture cards as well.
The minimum PPD considered acceptable is 2.4m or so for 6900xt.
Let me get home and see what it ran at during testing and will see if there was a mistake of not excluding this project from AMD or there is some issues with system at hand.
The reason we see AMD RX 6000 series lumped together as well as 5000 series, it's due to the way AMD shortcuts their way through PCI IDs, that cannot be changed unfortunately with current system. People from AMD involved with the project know about this, however they are too low in the food chain to change how AMD registers their GPUs.

[Starman157]

As for GCN? I have a 7950 that's GCN. Relatively speaking, it runs just the same as a NAVI. About 2.5x slower with this WU set.

Yes, I am now aware, by others, who've explained the shortcuts that AMD is using to identify their cards. If they are aware of this problem and choose to do nothing about it, must be for reasons other than card discrimination.

Marketing perhaps? If 6xxx cards are lumped together, it'll make the low end of the series "appear" to be better than they are as they get lumped in with the high end. Of course, the high end suffers because of it. I would imagine AMD sells far more low end cards than high end, so the possible marketing ploy could entice far more potential customers to the lower performing cards. They have however segregated the 6700, 6700xt and 6800M (mobile) into a different PCI Id (1002:73df) so at least these units won't possibly affect the 6800 to 6900xt "category". Or am I incorrect on this too based on FAH categorization?

Either that, or they're running out of hex numbers for product IDs (which I highly doubt).

[muziqaz]

OK,
Project ID: 17721
Credit: 12433

Proc: Radeon RX Vega 64

Min. Time / Frame : 00:00:56 - 1,130,201 PPD
Avg. Time / Frame : 00:00:57 - 1,100,590 PPD


Name: Ryzen 5950x
GPU / Radeon RX 6900 XT

Min. Time / Frame : 00:00:35 - 2,287,367 PPD
Avg. Time / Frame : 00:00:35 - 2,287,367 PPD

As this project only has 92k Atoms, it is relatively small. 2.2m PPD for 6900xt is at the bottom of what would be acceptable.
1.6m seems to suggest some bottlenecks or other interference on your system

During testing projects are run through Vega 64, Radeon 7, RX5700xt and 6900xt for AMD GPUs. These tend to give a good indication which projects run ok on AMD which will not.
Unfortunately, 5700xt and Radeon 7 did not manage to grab one of these to run.

Reasoning behind why AMD does what they do is beyond us, and if they are not changing their way, there is no reason to discuss it any further
PCI-ID is not really visible to customers, so it shouldn't influence any of buying decisions anyways. Those IDs usually are visible to devs (drivers, Linux patches, etc)

[Starman157]

Nope. My system falls into the averages listed here.

17714 - 1.467m https://folding.lar.systems/projects/fo ... file/17714 RX6xxx series is ranked 18th, RX Vega 56/64 is ranked 30th at 640,165
17716 - 1.455m https://folding.lar.systems/projects/fo ... file/17716 RX6xxx series is ranked 18th, RX Vega 56/64 is ranked 34th at 412,865
17717 - 1.738m https://folding.lar.systems/projects/fo ... file/17717 RX6xxx series is ranked 15th, RX Vega 56/64 is ranked 29th at 396,586
17720 - 1.697m https://folding.lar.systems/projects/fo ... file/17720 RX6xxx series is ranked 16th, RX Vega 56/64 is ranked 28th at 610,316
17721 - 1.721m https://folding.lar.systems/projects/fo ... file/17721 RX6xxx series is ranked 17th, RX Vega 56/64 is ranked 32nd at 432,878
17722 - 1.840m https://folding.lar.systems/projects/fo ... file/17722 RX6xxx series is ranked 18th, RX Vega 56/64 is ranked 28th at 722,548
17723 - 1.991m https://folding.lar.systems/projects/fo ... file/17723 RX6xxx series is ranked 14th, RX Vega 56/64 is ranked 30th at 297,731
......
17742 - 1.741m https://folding.lar.systems/projects/fo ... file/17742 RX6xxx series ranked 14th, RX Vega 56/64 is ranked 27th at 106,550 (ouch)

You get the idea.

I guess what you tested is quite a bit different than what is out in the wild running on other machines.

[muziqaz]

ok, give me TPF values instead of 3rd party database where it only has listed total folding time of the WU. I have no clue how that system is calculating PPD, what I do know is TPF values from which we calculate PPD. If you start calculating your PPD from the moment you began folding to the moment of the last bit going back to server, you will get a inaccurate (however realistic) PPD values. Why inaccurate? because every single person's internet connection speed, and connection between folder and servers are different. We do not take that in consideration because upload time is a variable, not a constant. TPF is constant in a sense that it does not varies.
All these listed projects show minimum 2.6m PPD for 6900xt, and are you sure that 6000 series entry in that database is not 6800? No, you are not.
That lar system should be used as approximate information, not an exact calculation of PPD values

The only way this conversation will continue, when you give me your 6900xt and your Vega 64 TPF values for these project. Then we can compare apples to apples instead or tomatoes to oranges

[Starman157]

ok, give me TPF values instead of 3rd party database where it only has listed total folding time of the WU. I have no clue how that system is calculating PPD, what I do know is TPF values from which we calculate PPD. If you start calculating your PPD from the moment you began folding to the moment of the last bit going back to server, you will get a inaccurate (however realistic) PPD values. Why inaccurate? because every single person's internet connection speed, and connection between folder and servers are different. We do not take that in consideration because upload time is a variable, not a constant. TPF is constant in a sense that it does not varies.
All these listed projects show minimum 2.6m PPD for 6900xt, and are you sure that 6000 series entry in that database is not 6800? No, you are not.
That lar system should be used as approximate information, not an exact calculation of PPD values

The only way this conversation will continue, when you give me your 6900xt and your Vega 64 TPF values for these project. Then we can compare apples to apples instead or tomatoes to oranges

I believe that LAR systems are using API calls to get the PPD values from the FAH Client directly. So whatever the client calculates is what is recorded and shown, and have seen visually both FAH Control and the Web Control clients indicated the same values. No idea if LAR systems is recording TPF Values along with information returned back to them; it isn't mentioned.

Yes, I've seen the recorded PPD value drop from the finished value as it's recorded in the log when the actual WU is uploaded. So the discrepancy for all these "averages" I guess must be attributed to internet connection and server delay. OK, inaccurate, but realistic. It must take some serious internet connection and server delay to cause all 6xxx GPUs to report the averages indicated. Seems only the AMD products are affected. Interesting. I am now wondering how long the internet and server delay must be to cause PPD values to drop from your stated 2.6m to the numbers realistically reported at LAR systems.

So all the minimums for the 6900xt are 2.6? I guess it isn't all then as you've invalidated the acme as you calculated 2,287,367 PPD and a TPF of 35 seconds? Also, ALL the minimums for the 6900xt, when there's no stated discrimination in the 6xxx category for any of the 6900xt? Again interesting. As for the 6xxx entry in that database only being 6800? How is that possible, again, without FAH discrimination of the 6xxx series. I believe what is reported at LAR systems is the exact same "bucket" that FAH uses, namely "RADEON RX 6800/6800 XT / 6900 XT". As for the composition of this bucket as to the population of 6800s, 6800xts and 6900xts, based on price, I would say the former have a higher proportion than the latter. A 6800xt is not that much slower than a 6900xt. Same for the 6800.

But all this started with a comparison about 6900xt to 3080. If all things are equal, with the 6900xt being "slower" by about 22% than the 3080 (as stated by Bruce) in FP32, the 3080s realistic reported value of 6,011,894 PPD on WU 17721 should equate to something like 4,689,277 for the 6900xt. It doesn't. The present realistic average reported value is 1,713,874. That's not a drop of 22%, but 71.5%. That must be some very serious internet and server delay to cause that. Again, it appears to only affect AMD systems.

As for the slowest of the category, the 6800, is 45% slower than the 3080. Realistic reported values should be around 3,306,541. The 6800XT realistic reported value should be somewhere around 4,208,325. Nowhere could I get any of the 6xxx contingent to calculate out so that any one of the members could sway the average down to 1,713,874.

So you asked for a TPF. My 6900xt is working on WU 17717 (prcg 82,3,73). FAHControl 7.6.21 reported values are: Estimated TPF: 37 seconds, Estimated PPD: 1,615,234. Funny how the TPF is very close to your 35 seconds, but the Estimated PPD is nowhere close to the 2,287,367 you reported. Or should that be 2.6m as you say ALL the 6900xt WUs in this project are above that number?

Am I to surmise that even the paltry 1,615,234 value will drop significantly with internet and server delays?

All I'm saying is that there's a huge discrepancy between the AMD PPD values and the Nvidia PPD values and it appears to affect the project 17711 to 17743.

It's also interesting that the WU right after the 17717 above, my 6900xt is working on WU 17340. FAHControl Estimated PPD is 5,202,940 (highest) to 4,521,235 (lowest). TPF: 40 seconds to 44 seconds. Based on 22% slower than a 3080 (with realistic reported values of 5,477,855 courtesy of LAR Systems) it would appear my 6900xt is WAY overachieving its 4,272,726 supposed PPD value based on FP32 relative performance. Or is the 3080 underperforming?

Seems I've come to the conclusion that "all things AREN'T equal" adjusting for relative performance, even though there is the illusion that they are.

If you are trying to achieve performance parity on reported PPD values, then I thought it best to report it to you.

BTW, If you are waiting for Vega values, I don't have that GPU. You reported on the Vega, so I thought it apropos to include their realistic reported values as an indication that it's not just a 6xxx issue.

[muziqaz]

We can sit here for days and guess what LAR system is recording.
Instead I will tell you what we use to calculate PPD, and what I have as evidence that 6900xt are not getting 1.6m PPD using TPF method
In real world 6900xt is equal to 3090, 6800xt is equal to 3080. In folding, this is not the case, and never was. PLEASE stop using nVidia alternatives as your argument. Do you want, that I suggest to researcher to disable AMD cards completely, instead of them helping science somewhat?
Until AMD starts involving themselves with FAH project to the same level as nVidia is, you will NEVER get same performance from AMD cards compared to nVidia. PERIOD.
This topic has been beaten to death many many times (by myself, of all the people). AMD does not want to get involved (believe me we asked), nVidia however has some very dedicated people. If nVidia did not have CUDA, AMD would have somewhat caught up with them, however recently introduced CUDA optimisations pushed nVidia cards further up.
Now, instead of picking on 2.6m being a set minimum, let's concentrate on other more pressing matters, like Easter weekend

P.S. How about anything above 2m for 6900xt is acceptable You DO realise that if we increase base credit as per your desire, that will translate to even higher PPD for nVidia cards as well. And if you don't want AMD high end cards being assigned smaller projects, the switch is: All AMD cards, or none, there is no in between