Hardware incompatibility

[kriesel]

This is the place for sagas of driver updates gone wrong, Hatfield & McCoy type experiences with feuding graphics card drivers, etc.


Share your best battle story here!

[kriesel]

Bought a used computer, which had an old NVIDIA NVS295 card in it, and added some RX550s to run gpuowl and mfakto. The NVS295 was capable of running a display and measured only a couple GhzD/day in mfaktc so not worth running that. Well. That was the plan. The mixed system turned out to be so much trouble, that I ended up removing the NVIDIA card and putting it in last-resort-spare-parts storage, running the NVIDIA uninstaller, and rooting all through the file system and registry manually to get rid of what the uninstaller failed to clean out. A couple rounds of that interspersed with restarts of the Win7 system finally allowed the NVIDIA traces to all be removed, and AMD use become stable.

[kriesel]

Started a mini-miner rig build, with its onboard IGP, i7-4790, and 6 pcie slots, with an old HD with Win7 on it. Fine. Added an idle RX550, still fine. Mfakto and gpuowl run on the RX550.
Had some old CC2 NVIDIA cards, added them, still fine, could run mfaktc and cudalucas on them too. Added a CC3 NVIDIA card, still fine. Added a nice fast new RTX2080Super for some horsepower. This caused an NVIDIA driver to a version that supports CUDA10.x, which disabled all the CC2 gpus. AMD still working. Somewhere along the way, an Asrock Radeon VII joined the party. It oddly required an old AMD driver version, 19.5.1. Still fine. Removed the CC2 NVIDIA cards. Added a GTX1650, and suddenly the Radeon VII could no longer be used.
During this sequence I had noticed that GPU-Z would periodically indicate OpenCL active on the HD4600 IGP, but it wouldn't stay long enough to be useful.
Added another GTX1650. This one is unusual in that it has an auxiliary power jack and an 85W rating; most are 75W and powered by the PCIe slot.
As a test, disconnected the NVIDIA gear, and found I could run the RadeonVII again.
Reactivate the NVIDIA gpus, and the RadeonVII disappears again.

Ok, try a clean reinstall of graphics drivers.
Power down, disconnect all AMD, NVIDIA are connected, restart, reinstall NVIDIA graphics drivers. Should put the NVIDIA house in order.
Power down, disconnect NVIDIA, reconnect AMD, attempt to reinstall AMD driver, can't complete. Attempt clean removal, can't, because Windows Update is running well hidden. Get rid of Windows Update by repeated updates and restarts and task manager terminations, finally get a clean removal of AMD. Restart. Reinstall the 19.5.1 driver AGAIN. A couple more restart rounds and it's capable of running on both the RX550 and Radeon VII.
Power down, disconnect NVIDIA and AMD, remove Intel graphics driver, restart. Reinstall Intel graphics driver. Restart. STILL, no OpenCL on IGP function.
Give up on that minor throughput possibility.

Shut down and reconnect NVIDIA and AMD, restart and wait. Text crash dump on IGP-attached monitor. Dump files being saved. Then loading system repair stuff. Something about failed to start Windows in order to prevent damage to the hardware. It was unable to effect a system repair, and shut itself down. I removed the Radeon VII, and all's fine again; I can run the RX550 and all the NVIDIA. Which were from the gpuowl point of view, in ascending device number order, RTX2080Super, GTX1650, GTX1650, Quadro K4000, RX550, RadeonVII.

https://cryptovoid.net/amd-nvidia-in-same-machine/ seems similar, and advises to install AMD first, then NVIDIA. https://forum.ethereum.org/discussio...-and-if-so-how advises NVIDIA first then AMD. Haven't found much about getting the IGP to work in OpenCL. https://stackoverflow.com/questions/...a-installation is close but is about cpu. I've had gpuowl -h showing NVIDIA, cpu, and AMD together, but not the HD4600 IGP. Minor throughput so low priority.

So I'm now at the point where I'm ready to resegregate (gpus). Diversity is too much trouble. More NVIDIA will be relocated to this system, and the AMD to the system that the relocated NVIDIA vacated from.

[tServo]

A few months ago, I wanted to upgrade a older machine that had a modest AMD video board with a more powerful Nvidia board I wasn't using so I could run GPUOWL on it.
After the gpu swap, everything was fine until I actually tried running GPUOWL. It complained that it couldn't find OpenCL ! This was in spite of the fact that I uninstalled the AMD drivers, swapped boards, then installed new Nvidia drivers.

I then did a scan looking for opencl.dll and found many AMD copies lurking all over the place. After trying to replace them with the Nvidia dll, it still failed.
As Gru would say: "Light Bulb!"
I installed the full Cuda toolkit and that fixed the problem.
All I can guess is that the Cuda toolkit plopped Nvidia's opencll.dll in a place that the video driver didn't.

BTW, as indicated by messing around with DLLs, this is a windoze box.

[kriesel]

Round 2 part 2
Removed the RX550 to spare status, and Asrock Radeon VII which had failed early (RMA for refund), making the mini-miner NVIDIA-only, leaving two unused slots on the mini-miner.
My nature abhors a vacancy. So I acquired an RTX2080 and a GTX1650 to fully populate the mini-miner. (MOAR POWR!)
The RTX2080 and RTX2080Super appear to require mutually exclusive driver versions. So, I thought, swap the RTX2080Super with a GTX1080Ti in a third box.
The Quadro K4000 got kicked out by the RTX2080 requiring NVIDIA driver 445 which is CUDA 11 but apparently not CC 3 compatible. Or maybe it's some other type interaction, but it's only ~110GhzD/day but 120W so not a priority especially with warmer weather coming.
So that makes the mini-miner all NVIDIA, but it seems like every gpu is a different brand. The third GTX1650 doesn't get along with the GTX1080Ti; only one is recognized if both are installed. Ah, but there's power and space for it in box 3.
Nope, no such luck, it will launch an app there but goes awry within seconds (error cudaGetLastError() returned 702: the launch timed out and was terminated), too soon to get anything done, recovery takes a system restart or gpu disable/enable cycle in device manager, and repeat. Regedt32 to increase TdrDelay has not resolved it. Some say it is a motherboard problem, or a driver version problem. https://render.otoy.com/forum/viewto...p?f=30&t=63360
Meanwhile the mini-miner has decided, it will no longer accept the presence of the backup HD, refusing to boot if it is connected, going into BIOS setup or System Repair instead.
There's plenty of power margin; it's drawing 720W at the wall, 1200W rated.
Maybe displace GTX1080 from box 4 which overheats its room and trips that system's power limit if TF is run, over to the mini-miner. But that will also dislodge another Quadro 2000 which is only CC2.1 and running doublechecks...
It's become a sort of whackamole, or union bumping cascade.

[LaurV]

Wait! What?
You had Radeons and Geforces mixed in a rig, and it worked? My former understanding was that they bite each-other. If so, then should I put back the forzes together with the sevens, or at least the two big-irons, and try to tame them together? (I have 7 slots there, and the mobo used to run 6 Titans classic and one black, in her prime age... but now it would be a bit... "crowdy" with some of the raisers removed and lost/damaged some time ago)

[kriesel]

Originally Posted by LaurV

Wait! What?
You had Radeons and Geforces mixed in a rig, and it worked?

In a fragile way and temporarily. The Asrock Radeon VII was determined to have failed early, by trying it in 3 systems and observing it prevented boot. RMA for refund within 30 days. Proceed with caution if at all. I've done it twice, but subsequently gone AMD-only or NVIDIA-only per system.

Some miner motherboards require mixing NVIDIA and AMD, since they have more pcie slots than an individual vendor's drivers will support gpu count. I think those are usually run with linux. https://www.asus.com/support/FAQ/1034921
Financing and cooling for such a beast could be a challenge.
https://www.aurelp.com/tag/how-many-...ws-10-support/
If attempting such a build, I would go with all same model same brand NVIDIA and AMD respectively. Mixing NVIDIA models and brands has also proven problematic.