Add Vega 8 Mobile Graphics

[Smith6612]

Hello there,

I recently installed the Folding@Home client to a new laptop I've recently purchased, and wanted to get the integrated Vega 8 Graphics added to the GPUs.txt whitelist. Currently, CPU folding is working fine. Attached is the GPU-Z information.

[bruce]

We have very little experience with the Ryzen family of GPUs. All laptop GPUs are designed to be low-power devices and they may or may not be fast enough to process FAH assignments within the deadline. I will enable your device, but you'll need to give me feedback about whether it can meet the deadlines. If not, there's no point in assigning WUs to folks with that hardware.

[Smith6612]

Thank you Bruce.

I am seeing some strange behavior right now with the stability of the software, after adding the GPU Slot. I had a case shortly after enabling the GPU slot where the Core_a4 CPU slot AND the GPU Core 21 slot crashed out at the same time. CPU remains stable so long as GPU isn't doing anything. I had a crash in the GPU driver which resulted in a bad work unit submission to occur, so it may be safe to say at the moment the Vega 8 OpenCL drivers are not exactly mature. But with that said, the compute aspect does work once a work unit has begun.

On some good news, it appears the Time per Fold for the Work Unit I got is approximately 6-9 minutes. The chip appears to be promising from a compute standpoint.

[bruce]

I am seeing some strange behavior right now with the stability of the software, after adding the GPU Slot. I had a case shortly after enabling the GPU slot where the Core_a4 CPU slot AND the GPU Core 21 slot crashed out at the same time. CPU remains stable so long as GPU isn't doing anything. I had a crash in the GPU driver which resulted in a bad work unit submission to occur, so it may be safe to say at the moment the Vega 8 OpenCL drivers are not exactly mature. But with that said, the compute aspect does work once a work unit has begun.

On some good news, it appears the Time per Fold for the Work Unit I got is approximately 6-9 minutes. The chip appears to be promising from a compute standpoint.

If you the chip is promising from a compute standpoint, then enabling it was the right thing to do.

The embedded GPU shares chip real-estate and a heat-sink with the CPU. Running them both at full-load produces a maximum amount of heat and the installation probably wasn't built to handle it all, hence the simultaneous crash. I'd say you probably have to choose one or the other and let the other one operate at a minimum load level. Adding more fan power might be possible but my guess is it won't be feasible.

[bruce]

The "power" slide does adjust the number of CPUs allocated provided you use the default (-1) setting.

Never use the cpu:7 setting, whether you set it ore the system sets it.

You can use 1,2,3,4,5 or 6, (or none) depending on what works. Even if you decide not to fold with your CPU, try to FINISH the current WU. (CPU:2 might be a good place to start.) FINISH will avoid downloading a new WU but only until you restart the client, so if you decide on zero, you'll need to remove the CPU slot once FINISH has done its job.

[Smith6612]

Thank you.

There is definitely some driver instability, as a Windows crash occurred about a half hour ago. The work unit itself continues to move onwards. As for the CPU slot, the folding client selected the value of 7. I'll decrease it to 6 or 4 depending on how things go with the GPU slot.

I found that Windows Update has a slightly newer GPU driver in it (although for some reason the version number is .1 lower despite the newer build date?) than what HP's website offers. I am going to give this driver a try for a little while.

As for temperatures, the GPU and CPU under full load are getting no hotter than 70C, and I am seeing no evidence of thermal throttling. Because Vega 8 is so new, tools like GPU-Z don't read the clock speed correctly from the chip. The Ryzen component is the only side I can fully validate. It's possible the crash is power related too, which I will test out if I see any further crashes.

[JimboPalmer]

I often find that MS updates do not work with F@H, if so, reinstall the HP driver.

[bruce]

Like JimboPalmser says -- When Microsoft updates the dirver for you, it REMOVES OpenCL support so it breaks FAH. I normally get the current version directly from NVidia which prevents Windows10 from removing OpenCL.

Sometimes laptops run special driver versions, so I can't promise NVidia's drivers will successfully replace the ones from HP. -- but my guess is that's likely to be a good plan.

[foldy]

Microsoft nvidia driver for Windows 10 included OpenCL lately but I don't know about AMD.

[Smith6612]

It appears in this case, the driver from Windows Update is more stable with Folding@Home than the one offered by HP's website. Still seeing stability issues, and I'll track them down once I have some free time. I have a feeling the crashes remain specific to certain types of workloads related to compute. A work unit was able to be completed successfully overnight by the GPU, which is good.

[_r2w_ben]

Please let us know when you have some PPD numbers.

I have a Kaveri A8 7600 3.1GHz (3.8GHz boost). When folding on the GPU, the CPU down clocks to 2.4GHz to stay within 65W. Other tasks can cause the CPU to boost but then power usage briefly goes over 80W.

CPU slot 2 threads
1500PPD on Core A4 work units
3200PPD on Core A7 work units

GPU
12-18K PPD

Unlike 1080TIs and Tesla V, this produces more PPD on p94xx and less on p114xx work units.

[foldy]

Maybe it is possible to increase the power limit in bios or using other tool?

[bruce]

Unlike 1080TIs and Tesla V, this produces more PPD on p94xx and less on p114xx work units.

I have not studied this carefully, but if somebody wants to, look for other projects to add to those two groups. I suspect there is some kind of correlation between the protein's total atom count and whether they work better on group p94xx or group p114xx.

There may be something else involved, but without some additional reports, I don't have enough data to adequately describe what the OpenMM experts should investigate. I know there's a problem, but I don't know how to describe it in detail. Without a detailed problem description, it'll be treated as a low-priority issue since "everything works."

[Smith6612]

Just a small update. I resumed folding onto the GPU. The stability issues with the drivers are a strange one. I found that if I have the GPU continuously decode video, such as from Twitch.tv or YouTube, the driver NEVER crashes even with Folding@Home completing work units. It seems the GPU can complete 2-3 units a day on Core 21 at that rate.

Such a weird driver problem, though! Who would've thought that putting more load, let alone a more diverse load onto the GPU would make it more stable? More surprising, the chip seems to have a lot of headroom still, as Folding@Home and video decoding don't slow down the UI in Windows at all.

I probably will not run Folding@Home on both the CPU and GPU until I can either get a fixed GPU Driver, or until a UEFI update is released with some updated microcode for the system.

[rwh202]

Just a small update. I resumed folding onto the GPU. The stability issues with the drivers are a strange one. I found that if I have the GPU continuously decode video, such as from Twitch.tv or YouTube, the driver NEVER crashes even with Folding@Home completing work units. It seems the GPU can complete 2-3 units a day on Core 21 at that rate.

Such a weird driver problem, though! Who would've thought that putting more load, let alone a more diverse load onto the GPU would make it more stable? More surprising, the chip seems to have a lot of headroom still, as Folding@Home and video decoding don't slow down the UI in Windows at all.

I probably will not run Folding@Home on both the CPU and GPU until I can either get a fixed GPU Driver, or until a UEFI update is released with some updated microcode for the system.

In 10 years of folding, I've never had a truly stable AMD GPU - normally crashes would correspond with changes in power state - voltage seems to be dropped fractionally too soon. By keeping load on, there is never a need to drop to a lower power state and so better stability.