Any way to get Navi (Radeon 6xxx) GPUs to perform better?

[psaam0001]

This is one of AMD's self-inflicted flaws: GPU driver's. NVidia updates theirs more frequently; where as AMD is putting their efforts on making their hardware (the GPU chips) w/o focusing on what allows that hardware to interact with the O/S (drivers that run smoothly).

Paul

[jchang6]

I just switched my 5500 (Win10) and 5600 (Win11) system to 21.3.2. It does seem to be faster based of last run on 21.10.1 vs current run on 21.3.2. Its too bad we don't a record of project + credit + run time for same project comparisons

[aetch]

They're called logs.
They record just about everything from project, start and finish times, to estimated credit (this s usually within about 20 points of awarded points).
They record your system configuration when the FAH client was started, including driver versions, and changes you have made to the folding slots on-the-fly (adding/removing a CPU/GPU slot, changing the size of the CPU slot).

Granted, there are some changes they don't log like cpu/gpu tweaks done with other tools (e.g. MSI afterburner, Windows processor power management) to increase/lower power usage by increasing/lowering clock rates, but that's a whole other rabbit hole.

[muziqaz]

You need to use 21.3.2 driver for your PPD to be somewhat decent (AMD in general is crap at folding). Any later drivers half the PPD. I reported this to AMD half a year ago, they are yet to acknowledge this issue in their release notes.
I would urge to report this issue to AMD GPU driver team through their driver suite Bug report utility. The more of us shout at them the more chance someone will wake up. Not the best behavior from so called project partner. smh

Keep in mind 6900xt gets as low as 2.6m ppd in majority of the projects even with 21.3.2 driver. On the upswing, max I have seen is 4.8m PPD I believe. So adjust that to your 6600xt. nVidia's 3000 series in CUDA mode still miles ahead

Thanks for the advice, but 21.3.2 doesn't support the 6600xt - it wasn't released yet.

damn it, forgot about that
so only solution (besides hacking the older drivers): spam all available AMD social media feeds and their bug report utilities with these issues

[muziqaz]

You need to use 21.3.2 driver for your PPD to be somewhat decent (AMD in general is crap at folding). Any later drivers half the PPD. I reported this to AMD half a year ago, they are yet to acknowledge this issue in their release notes.
I would urge to report this issue to AMD GPU driver team through their driver suite Bug report utility. The more of us shout at them the more chance someone will wake up. Not the best behavior from so called project partner. smh

Keep in mind 6900xt gets as low as 2.6m ppd in majority of the projects even with 21.3.2 driver. On the upswing, max I have seen is 4.8m PPD I believe. So adjust that to your 6600xt. nVidia's 3000 series in CUDA mode still miles ahead

I wonder if AMD (Lisa Su) knows how much they're missing out on people folding not buying their GPUs.
Not that it matters, because AMD GPUs always have less shaders than Nvidia of the supposed same category.
It's like a 6600XT should be the same as a 3060, in some games, but Nvidia GPU has double the shaders/cuda cores at only about 20% lower core frequency; resulting in much more efficient folding.

nVidia has a man who is co founder of folding@home project, so his dedication to optimize and push performance paid off. AMD's involvement with the project is more shallow. The middle management knows the issue, but it is low priority for them unfortunately. If Lisa Su knew about this, I am sure we would have person or two optimizing things. I ran out of voice and favors by now trying to pull someone from AMD to get involved...

[XanderF]

Not that it matters, because AMD GPUs always have less shaders than Nvidia of the supposed same category.
It's like a 6600XT should be the same as a 3060, in some games, but Nvidia GPU has double the shaders/cuda cores at only about 20% lower core frequency; resulting in much more efficient folding.

Eh, not quite double, and not 20%. But the bigger difference is in how those cores are structured.

Code: Select all

GPU				Cores	Clock 				TDP					Approx Gflops Processing Power (Base-Boost)
						     (Base-Boost)				    Half Precision		Single Precision		Double Precision
Geforce 3060 	3584	1320-1777			170w		9462-12738			 9462-12738				148-199
Radeon 6600xt	2048	1968-2589			160w		16122-21209			8061-10605				504-663

For single-precision calculations, the 6600xt does trail the RTX 3060. OTOH, for half-precision or double-precision calculations, the 6600xt is CONSIDERABLY more capable. Either way, the 3060 raw numbers don't explain how the card is getting triple the PPD of the 6600xt - these cards SHOULD be trading blows, more or less.

[XanderF]

so only solution (besides hacking the older drivers): spam all available AMD social media feeds and their bug report utilities with these issues

I was about to question this, but sure enough the Radeon driver interface does have 'poor performance' listed as an option in the 'report a bug', so...works for me! If AMD considers poor performance a bug, I'm happy to file a bug report about it...

I did include links to the folding.lar.systems site for comparison, to show this problem isn't specific to my system, but more broad.

[Neil-B]

There may also be something else to be factored in,

Many/most Projects have WUs that will maximise usage of thinner/slower GPUs but struggle to fully utilise faster/wider ones. When Project WUs are benchmarked the BC for the WUs will accurately represent/reward the performance of the slower GPUs but since the faster/wider ones are actually not folding as fast as they could then the ppd which represents the amount of folding done not what could have been done with fast/wide optimised WUs is used. nVidia GPUs have cuda which can make a big difference especially on the faster/wider GPUs and make up for some of the low ppd - but it is only on projects such as 18201 that the faster/wider cards truly show their form.

LAR shows the average of what has been folded and as such probably shows ppds for faster/wider gpus that are considerably lower that they would be if the WUs actually fully used these GPUs ... Obviously for slower/thinner GPUs the figures are more reliable as even a small wu can fully utilise these.

AMD gpus simply from not having cuda may be 25-40% lower than an equivalent nVidia one

[muziqaz]

so only solution (besides hacking the older drivers): spam all available AMD social media feeds and their bug report utilities with these issues

I was about to question this, but sure enough the Radeon driver interface does have 'poor performance' listed as an option in the 'report a bug', so...works for me! If AMD considers poor performance a bug, I'm happy to file a bug report about it...

I did include links to the folding.lar.systems site for comparison, to show this problem isn't specific to my system, but more broad.

There is no need to sugar coat their ignorance with poor performance option. They need to be blasted big time for this. Not only have they been completely ignoring f@h optimization wise for year, but they are also ignoring the fact that they dropped performance specifically for f@h by ~50% with their newer drivers.

[MeeLee]

Not that it matters, because AMD GPUs always have less shaders than Nvidia of the supposed same category.
It's like a 6600XT should be the same as a 3060, in some games, but Nvidia GPU has double the shaders/cuda cores at only about 20% lower core frequency; resulting in much more efficient folding.

Eh, not quite double, and not 20%. But the bigger difference is in how those cores are structured.

Code: Select all

GPU				Cores	Clock 				TDP					Approx Gflops Processing Power (Base-Boost)
						     (Base-Boost)				    Half Precision		Single Precision		Double Precision
Geforce 3060 	3584	1320-1777			170w		9462-12738			 9462-12738				148-199
Radeon 6600xt	2048	1968-2589			160w		16122-21209			8061-10605				504-663

For single-precision calculations, the 6600xt does trail the RTX 3060. OTOH, for half-precision or double-precision calculations, the 6600xt is CONSIDERABLY more capable. Either way, the 3060 raw numbers don't explain how the card is getting triple the PPD of the 6600xt - these cards SHOULD be trading blows, more or less.

The core frequency of Nvidia GPUs is kind of BS. I have never owned a (modern RTX) Nvidia GPU that couldn't hit 1850Mhz constantly, and 2035Mhz under high boost/overclock.
The lowest I set the core frequency is ~1835-1877Mhz depending on the load. Any lower, and it doesn't make any sense.
Not sure how much AMD hits, but most examples I've seen, run at 2-2.2Ghz.