AMD RX6000+ family GPU owners, please update to 22.7.1 drivers for optimal F@H performance

[muziqaz]

Hi,

as cs9k already reported in another thread, AMD finally fixed OpenCL performance with their latest driver version (22.7.1).
Until today majority of RX6000 series GPUs were folding at ~50% of it's capabilities and the last driver which was still offering good performance was 21.3.2, which was not compatible with majority of RX6000 series GPUs. Now your RDNA2 GPUs can match nVidia OpenCL performance with 22.7.1.

If you have older AMD GPUs, you can keep folding on 21.3.2 or older.

For those wondering how much you will gain? Your PPD pretty much will double, or even triple in certain scenarios.
If you are seeing 2.5-5m PPD with your 6800/6900xt GPUs, that means new driver is working for you.

Good luck, and thank you for your contribution

[toTOW]

I set this post as a Global announcement. Such long awaited big news deserve it ...

[cs9k]

Hear hear!

[psaam0001]

I saw the performance boost on my RX 6500 XT right away. Though things also may have stabilized as well on that Ryzen 3 3200G's iGPU as well.

Paul

[BobWilliams757]

I'm not sure if anyone else noticed, but in the release notes under Fixed Issues:
Lower than expected Folding@home™ compute performance with OpenCL™ API on some AMD Graphics Products such as the Radeon™ RX 6800.
Auto Undervolt may disable Zero RPM fan feature.


It might have taken them a while, but they did hear the concerns from the Folding@Home community. I've also seen people mention the zero RPM issue, so included that.

Though this issue seems to have impacted the faster cards the most, it looks like even my iGPU is going to see a boost based on initial WU's since the update.

[psaam0001]

The older Vega based iGPU cores that are in the 2000/3000 series Ryzen's may not see that much of a boost. But the RX 6000 series dGPU's will fare better with the right WU's.

Though, we will see what enhancements are in store from the F@H software/OpenMM development team(s), to optimize the AMD card performance in the future (as we saw with adding optimized CUDA support for the NVidia cards).

Paul

[psaam0001]

If you are having slow performance on an AMD iGPU while you are using a dGPU card, let the CPU task that you are running finish and then remove the CPU slot. Resources from the vacated CPU folding slot may be able to help on more complex WU's that are given to an iGPU (just a hypothesis--I need to monitor this for awhile to see if this helps for sure).

Paul (wishing there was a "CPU = False" expert setting)

[JimboPalmer]

To speed up my GTX 1060, I add the CPU slot option pause-on-start to true. I see no reason it should not work for AMD as well.

[muziqaz]

This is very dependent on project and a CPU. Yes, disabling CPU folding speeds up GPU folding, but in many cases speed up for AMD is very small.
If one wants to diversify their contributions, folding on both CPU and GPU has more benefits to science, than just on GPU

[BobWilliams757]

If you are having slow performance on an AMD iGPU while you are using a dGPU card, let the CPU task that you are running finish and then remove the CPU slot. Resources from the vacated CPU folding slot may be able to help on more complex WU's that are given to an iGPU (just a hypothesis--I need to monitor this for awhile to see if this helps for sure).

Paul (wishing there was a "CPU = False" expert setting)

It would be interesting to see if this works at all. I would suspect that if it does, then removing the CPU slot would help the iGPU as well as any dedicated card. It seems to me having it paused would have the same impact, but I've never looked at how much memory used/etc might vary with configurations. Then again, I've never yet bothered to allocate more memory to the iGPU to see if it impacts folding either.

I do know that I experienced changes over time with the various cores and how it impacted folding on both CPU and GPU at the same time. It seems that as cores advanced, it become more of a variable as to how much of a difference it made. Some WU combos would run the GPU at near normal TPF and the CPU side would suffer. Others would impact both or more the CPU. In at least one case the total PPD dropped below only folding one or the other, and I think that was the first time I had seen that.


These little iGPU chips are interesting in the way they allocate and share things, that's for sure.

[BobWilliams757]

OK... am I the only one on F@H that had issues with the new drivers? I got a couple crashes after installing them, and ruled out everything but the drivers before changing them. Reverted back, no issues.

So if nothing else, it's out here in the forums if anyone else has any problems. Ryzen 5 2400G, ASRock MB, quick memory.... and rock solid with F@H until the driver update. It only happened with heavy use (video editing) by a second logged in user on the system with folding in the background, and only a couple times. But that's a couple times more than it ever crashed in the past. Once going back to an older driver, I intentionally torture testing with F@H, a couple of benchmarks, and some web browsing and video editing all going on at once. The system didn't even blink.


On another note a question, since I haven't had this issue before. It was getting a driver timout error, but wouldn't reset on it's own. But F@H kept running the client, but couldn't work properly with the driver issue. So after the initial crash, it downloaded and dumped 4 or 5 work units before my wife filled me in that she had computer problems. Since this happened twice, it dumped 10 or 12 work units with errors. If this happened and wasn't detected, will the assignment server shut it down at some point and not issue another WU?

It was only taking 2-3 minutes to download, try, set and error code, and dump a work unit. I'd just hate to jump on the computer one day and find out it had dumped a whole bunch of work.

[PaulTV]

As for how the servers deal with work unit errors, as far as I understand... a certain error rate is accepted, and expected. If the error rate exceeds certain thresholds though, it will be flagged, and in extreme cases, job assignment may be halted. I don't think one could force such a situation from one machine though. But continuous job failure definitely is not something you want.

As for how those job errors may impact your own account: you receive quick return bonus points if you have successfully returned 10 jobs, and have a success rate of at least 80% overall. So if enough jobs fail, and the success rate drops below 80%... you get the idea. Btw, you can check your bonus status here: https://apps.foldingathome.org/bonus

[muziqaz]

OK... am I the only one on F@H that had issues with the new drivers? I got a couple crashes after installing them, and ruled out everything but the drivers before changing them. Reverted back, no issues.

So if nothing else, it's out here in the forums if anyone else has any problems. Ryzen 5 2400G, ASRock MB, quick memory.... and rock solid with F@H until the driver update. It only happened with heavy use (video editing) by a second logged in user on the system with folding in the background, and only a couple times. But that's a couple times more than it ever crashed in the past. Once going back to an older driver, I intentionally torture testing with F@H, a couple of benchmarks, and some web browsing and video editing all going on at once. The system didn't even blink.


On another note a question, since I haven't had this issue before. It was getting a driver timout error, but wouldn't reset on it's own. But F@H kept running the client, but couldn't work properly with the driver issue. So after the initial crash, it downloaded and dumped 4 or 5 work units before my wife filled me in that she had computer problems. Since this happened twice, it dumped 10 or 12 work units with errors. If this happened and wasn't detected, will the assignment server shut it down at some point and not issue another WU?

It was only taking 2-3 minutes to download, try, set and error code, and dump a work unit. I'd just hate to jump on the computer one day and find out it had dumped a whole bunch of work.

Updating to latest drivers with your set up is not the best idea. AMD and nVidia tend to "forget" older cards with their new drivers You can very easily fold and work stably with 21.3.2 driver.
On the other hand AMD was always shaky if one is folding on GPU and doing something else on same GPU

[BobWilliams757]

Thanks for the input.

The assignment issue had me wondering. It wouldn't take long to process a bunch of dumped work units if it wasn't noticed quickly. Hopefully it just never happens again, though I didn't think it would the first time really. The downside to F@H pushing hardware is you can't really run a proper stress test before the fact. And I've tried!

No worries about bonus status. Though it did bring my percentage of completion down, not nearly to the point of killing my bonus points.

As for the drivers, I just rolled the dice since I hadn't updated them in forever. I've never heard of the issues with AMD having issues if doing multiple things, and it really hasn't been the case on this system so far. It will happily chug away slowly even if CPU and GPU folding at the same time, with whatever my wife does (often some video editing) when I'm not on the computer. Since doing both types of folding on this slow iGPU doesn't help throughput much overall, I generally just do one at a time now, but use both as the occasional stress/fan/cooling test.

[Malique2]

Hi. Is there a reason my multi GPUs won't fold? All are AMD GPUs but they are showing active with no points accumulating.