Observation; 3D vs Compute GPU Workload setting

[NuovaApe]

This is not a problem report but a curious and beneficial (to me) observation.

Spoiler: my CPU is down from 85C to 50C and my GPU up from 85% to 99%, but the cooling fans aren't worried.

It centers around the Radeon setting "GPU Workload: 3D or Compute".

I lost my FAH virginity using CPU + GPU; i7 6700k + Radeon Vega 56, dial on Full. My PC was noisy as heck with the fans ramped to max by the various temperature sensors. My aluminium tower case was hot - worryingly so.

To help it cool down I slid off a case side panel. Much better. Lesson; get a CPU fan that helps transport hot air to the rear exhaust fan, rather than simply blowing hot air down onto the motherboard. I have a heat blender, not directed flow.

While the panel was off I installed my old Radeon RX480 alongside the Vega. I just had to know if it'd work.

It did. And 2 more fans to add to the din.

I deleted the CPU slot as the core temperature was now between amber/red. 2 GPUs is good enough.

I stumbled across Radeon Settings / Game Explorer / Global Settings / GPU Workload; 2 options - 3D or Compute. I switched to Compute and rebooted.

Something is wrong. Either I've gone deaf or FAH isn't running - the fans aren't full whack anymore.

FAHCore_21 is half the CPU of what it was (RX480 down from 10% to 5 and Vega down from 2% to 1). Task Manager shows GPU Compute at 99% continuous on both cards.

Given the lack of fan noise I added CPU back in via FAHControl. Previously CPU was flat out @ 50% (4 core, hyperthreading enabled) and the CPU fan was maxed out @ 2000RPM. Now it's 30%-40% CPU, almost as if sighing with relief.

I now have CPU + 2x GPU and I can no longer fry an egg on my aluminium case. All I did was switch from 3D -> Compute. What on Earth does it do under the hood? Nobody knows - I've googled.

In the i7+ the PCI lanes go through the CPU rather than North/South bridges of old (can't remember which). Could there be less PCI traffic now to warm it up? Could it be the vertical retrace interrupt is not needed for compute and was drawing many amps? I'm being facetious and showing my age.

Whatever the Compute setting does it is producing far less heat across my system, most notably CPU, and the work rate doesn't seem affected.

[X-Wing]

Test your GPU 3D performance in a good benchmark (3DMark, Unigine) under both settings, be sure to watch your temps and clock speeds closely, and see if the setting affects the performance of your cards in 3D workloads. Perhaps it's software optimization, or maybe it outright shuts off the 3D engine to save power and therefore heat. I'm honestly not sure, this is kinda a shot in the dark.

[bruce]

What on Earth does it do under the hood? Nobody knows -

Right

Go to the manufacturer's support forum and ask. If you get an answer, report back.

[NuovaApe]

What on Earth does it do under the hood? Nobody knows -

Right
Go to the manufacturer's support forum and ask. If you get an answer, report back.

I've looked on the usual; reddit, tomshardware etc. A guy asked on toms "what does Compute do?" and the mod was downright disparaging for the guy even wanting to know. A mod on an IT hardware focused website disparaging the quest to seek information on IT hardware

When I find out how Compute is implemented I'll let you know!

[MeeLee]

I would try to check Control, to see if you are still getting around 700k PPD for the Vega 56, and ~300k PPD for the Radeon RX 480.

3D in taskmanager, is basically just showing you the GPU usage to display the '2d desktop image' (unless you're watching youtube videos, or are simultaneously folding and gaming, then it should increase).
3D takes away from compute. If you see 3D increase, it usually goes paired with a compute decrease.
Compute should hover between 95-99%; but it doesn't accurately display what the GPU is doing.
So I wouldn't pay too much attention to the taskmanager, to determine what your cards are doing.

Since I don't own AMD GPU cards, I can only speculate,
It might be that the compute setting change, has introduced a bottleneck in your system.
Perhaps compute was made for bitcoin mining only, and the card isn't optimized for folding in this setting.
Whatever the case, check PPD range, and report back.
If performance differs by a lot from the above 700 and 300k PPD range per card, there's something wrong, and you should set it back to 3D.

If performance is about the same, there's a possibility your card has gone into low power mode, and disabled some of the unneeded parts of the GPU; like what we with Nvidia have to do manually.
If that's the case, the card is doing almost (within 90%) the same amount of work, at a much lower power consumption (could be between 10-50% lower consumption).
Sometimes, 20-30Watts lower power consumption, can bring the fans in a near 'noiseless' state; and gives the perception that the card isn't working at all; while in reality it's just using slightly less power.
I would measure the differences with a killawatt meter if you can, and report back.

[NuovaApe]

I would try to check Control, to see if you are still getting around 700k PPD for the Vega 56, and ~300k PPD for the Radeon RX 480.
3D in taskmanager, is basically just showing you the GPU usage to display the '2d desktop image'
I would measure the differences with a killawatt meter if you can

In Task Manager / Performance / GPU four of the stats boxes shown have a drop down. They default to 3D + Copy + Video Encode + Video Decode.

I switch them to "Compute 2" (Vega) and RX480 to "Compute 1" to see the activity. They correlate closely to the Radeon Wattman stats.

Additionally there's a bunch of GPU related performance counters in Windows perfmon (Performance Monitor). There's a ton of counters there. Assume TaskMan + Radeon are using that as the source.

As for PPD I've not been paying attention. I run FAH in spurts only when I'm home - PC is off when I'm at work.

PPD right now: CPU=21156, RX480=253558, Vega=545134. Credit=2063865 from 42 WU. If that's substandard I'm not too fussed. I like the quiet life and noisy fans do not help with that!

I do have an avo although it's a normal one not a killa one Can't be bothered to inline it to measure current. Whatever it is now is less than it was - my PC was unsustainable in 3D mode and I'm happy to let it trundle along in Compute mode. Besides, I pay the bills here

[bruce]

Ignore the PPD if you're not running 24x7. The PPD of any WU which doesn't run continuously will suffer because it depends on the projected upload time minus the download time.

Note the project number and the elapsed time for, say 10% progress on the WU. As long as you don't try to compare the rate of progress across multiple projects, it's a good measure of performance.
(Some projects are similar enough that you can compare 10% to 10% but you won't know if you happen to be looking at projects that are distinctly dissimilar.)

[NuovaApe]

I left my PC folding overnight and it completed a few more WUs contiguously:

PPD right now: CPU=24363, RX480=320032, Vega=806188

That's in line with what MeeLee said should be attainable, so it seems Compute mode isn't hurting performance.

I've a good quality "tower" fan arriving today to help with the poor cooling in my PC.

Sorry to hijack this thread but my post in "Problems with specific WUs" hasn't been answered - does it matter that I've been assigned projects that are "not found" via the project info page, eg
https://apps.foldingathome.org/project?p=13820

Thanks.

[MeeLee]

Those look more like the numbers you're supposed to be getting!
Now you can experiment with overclocking and power limiting, if you care to do so.

[Jone]

In 2017 the hashrate on AMD cards when mining ethash was falling as DAG file size was growing. The DAG file needs to be stored in the GPU memory. To remedy this AMD released a beta “blockchain driver” not intended for 3D rendering. After a while they introduced the compute switch in the regular driver, making the beta blockchain driver obsolete.
That is the story behind the compute switch. Exactly what it does is difficult to find out. Deals with GPU memory a bit differently?