New Radeon drivers - 21.10.1 - performance still seems bad

[XanderF]

As above - new drivers posted.

Release notes: https://community.amd.com/t5/knowledge- ... a-p/491713

Installed on my 6600xt, performance still seems as bad as any other recent release. Anyone else have different experiences with it?

[jchang6]

i did briefly, before going back to 21.3.2, but it will take time for project level stats

[XanderF]

i did briefly, before going back to 21.3.2, but it will take time for project level stats

Hmm. Curious if you get any more experience with it. Last couple days on 21.8.2, I was pulling 700-800k ppd on the 6600xt. Today with 21.10.1 I didn't manage better than 400k ppd over 2 different projects. Reverting to 21.8.2 for now.

EDIT: Although, TBH, maybe I shouldn't read too much into that. Looking at my logs, all I've gotten all day today was project 18110, which is godawful performance on Radeon cards, at least on any driver I can run. Just going to pull the GPU slot for now - not worth pouring 170w into ~300k PPD. Hell, I get the same PPD for less power using my CPU on its own. Like it's the 2000s again or something...

[jchang6]

here
viewtopic.php?f=38&t=37489
I did a break out of RTX performance by project
I have one machine with a RX5600 that than ran recent drivers, then I switched maybe a week ago? I will see if I can figure when each driver as applied to do a project breakdown

[XanderF]

I have one machine with a RX5600 that than ran recent drivers, then I switched maybe a week ago? I will see if I can figure when each driver as applied to do a project breakdown

Thanks - not that I think AMD is paying attention, but if we had some hard data comparing the same project on the same hardware, with 21.3.2 drivers vs the new ones - that would be solid evidence for them. Unfortunately, I can't help, as 21.3.2 doesn't support the 6600xt.

[muziqaz]

I have one machine with a RX5600 that than ran recent drivers, then I switched maybe a week ago? I will see if I can figure when each driver as applied to do a project breakdown

Thanks - not that I think AMD is paying attention, but if we had some hard data comparing the same project on the same hardware, with 21.3.2 drivers vs the new ones - that would be solid evidence for them. Unfortunately, I can't help, as 21.3.2 doesn't support the 6600xt.

The evidence has been provided, not that they need that, as they can easily run f@h project at the labs comparing with older drivers. Evidence or no evidence: AMD at the moment are not acknowledging issues with the project they claim they are partnered with.

[Neil-B]

EDIT: Although, TBH, maybe I shouldn't read too much into that. Looking at my logs, all I've gotten all day today was project 18110, which is godawful performance on Radeon cards, at least on any driver I can run. Just going to pull the GPU slot for now - not worth pouring 170w into ~300k PPD. Hell, I get the same PPD for less power using my CPU on its own. Like it's the 2000s again or something...

Researcher was made aware of the issue and has excluded AMD from this project.

[muziqaz]

EDIT: Although, TBH, maybe I shouldn't read too much into that. Looking at my logs, all I've gotten all day today was project 18110, which is godawful performance on Radeon cards, at least on any driver I can run. Just going to pull the GPU slot for now - not worth pouring 170w into ~300k PPD. Hell, I get the same PPD for less power using my CPU on its own. Like it's the 2000s again or something...

Researcher was made aware of the issue and has excluded AMD from this project.

This a shame, because project runs fine with 21.3.2 driver

Soon I hope we will completely exclude all AMD cards, just to make a point to AMD.

[Asghan]

Same problem here:
Every driver after 21.3.2 has a very poor OpenCL performance (6900XT here).
BOINC Projects are also slowed down.
As soon as I go back to 21.3.2 everything is fine.

I recommend to also report this via the feedback button in the driver menu to AMD.

[psaam0001]

If I wanted to place odds on whether or not I keep my AMD iGPU (on a Ryzen 3 3200G) folding, I would say there is an extremely good chance that (when my finances are better) a 5950X will be occupying that socket on the motherboard for this particular system.

Paul

[psaam0001]

OTOH: I might go ahead and be the guinea pig who tries the latest AMD drivers, after running DDU (in safe mode). Time for me to take the literal "jab in the shoulder", and test a work around.

Paul

[XanderF]

OTOH: I might go ahead and be the guinea pig who tries the latest AMD drivers, after running DDU (in safe mode). Time for me to take the literal "jab in the shoulder", and test a work around.

Paul

If you have a way to get good a-to-b numbers from it, that would be really interesting to see. Anecdotally, having now tried them off and on a couple times, they cut my numbers to 2/3 of the 21.8.2 drivers. (I cannot try the 21.3.1 at all for comparison - unsupported)

On a hunch, I tried copying over the 5 OpenCL .dll files from the 21.3.1 drivers into the 21.8.2 fileset, and 21.10.2 - in either case, the driver installs fine, with minor complaining from Windows about unapproved files, but both times OpenCL stops working in F @ H - "WARNING:FS01:No CUDA or OpenCL 1.2+ support detected for GPU slot 01: gpu:3:0 Navi 23 [Radeon RX 6600/6600 XT/6600M]. Disabling." And I'd also 'successfully' modified the 21.3.1 drivers to INSTALL in Windows for a 6600xt, but while it worked in safe mode, it couldn't render the fullscreen display (kept flickering on and off). So I think the bits or whole of 21.3.1 just isn't going to work on these newer cards.

If you have luck with an older card on the newer drivers, it'd definitely be interesting to hear a comparison point!

[psaam0001]

OTOH: I might go ahead and be the guinea pig who tries the latest AMD drivers, after running DDU (in safe mode). Time for me to take the literal "jab in the shoulder", and test a work around.

Paul

If you have a way to get good a-to-b numbers from it, that would be really interesting to see. Anecdotally, having now tried them off and on a couple times, they cut my numbers to 2/3 of the 21.8.2 drivers. (I cannot try the 21.3.1 at all for comparison - unsupported)

On a hunch, I tried copying over the 5 OpenCL .dll files from the 21.3.1 drivers into the 21.8.2 fileset, and 21.10.2 - in either case, the driver installs fine, with minor complaining from Windows about unapproved files, but both times OpenCL stops working in F @ H - "WARNING:FS01:No CUDA or OpenCL 1.2+ support detected for GPU slot 01: gpu:3:0 Navi 23 [Radeon RX 6600/6600 XT/6600M]. Disabling." And I'd also 'successfully' modified the 21.3.1 drivers to INSTALL in Windows for a 6600xt, but while it worked in safe mode, it couldn't render the fullscreen display (kept flickering on and off). So I think the bits or whole of 21.3.1 just isn't going to work on these newer cards.

If you have luck with an older card on the newer drivers, it'd definitely be interesting to hear a comparison point!

I'm going to link to a video on the JayzTwoCents YouTube Channel that will give more information about DDU, and how to use it in Windows 10 (in safe mode--before installing known working display drivers): https://www.youtube.com/watch?v=F8OLhUAPDq0

The only AMD video device I have that is capable of folding is the Radeon Vega 8 iGPU that is part of the Ryzen 2 3200G CPU.

Paul

[psaam0001]

Follow up: It got bad enough where I removed the updated WHQL drivers for both AMD and NVidia, and allowed Win10 to locate optimized drivers for the installed devices.

So.... Looks like I'm going to wait for the finances to improve, and plan for upgrades.

Paul

[Asghan]

Same problem here:
Every driver after 21.3.2 has a very poor OpenCL performance (6900XT here).
BOINC Projects are also slowed down.
As soon as I go back to 21.3.2 everything is fine.

I recommend to also report this via the feedback button in the driver menu to AMD.

is there any offical person from F@H who can reach out to AMD about this?
We are litteraly wasting performance and energy if someone is just always using the latest driver due to the reduced OpenCL performance.