When NVIDIA and Radeon drivers don't play well together for BOINC, I had to pick one or the other and get rid of the temptation to try to run both under BOINC.
I run with a dual AMD/NVIDIA setup most of the time, the biggest issue i have found is that both brands during the installation write over the same OpenCL file that causes issues with the othjer.
Hmm, are you talking about Windows? That's ironic because I have been fighting with the Linux AMD driver installer for approximately 20 years and, just recently, for my new 6800XT, I discovered that *no drivers* are required to make it work for BOINC. The ancient way of offering OpenCL support/APIs by overwriting libOpenCL.so, as well as the newer way of installing an ICD that allows for multiple OpenCL paths to exist simultaneously--which I assumed was how Windows did it, too--are *gone* because AMD has ended "direct" OpenCL support through their drivers and pushed all OpenCL support on top of ROCm. So, OpenCL doesn't talk directly to the AMD driver anymore. I think it goes kernel/driver/Vulkan/ROCm/OpenCL, now. If you have an RDNA2 card, you shouldn't be installing any OpenCL "drivers". You can even see it in the documentation for ROCm; it describes how the stack has changed with ROCm. The thought that, in Windows, they are still overwriting shared libs, blows my mind. That's a hack from before nVIDIA sold their first card.
Here is mystery I uncovered that proved it: Uninstall all the AMD drivers and install (or leave installed) ROCm, then run whatever program shows you OpenCL supported devices. My card showed up without any divers install. I couldn't believe my eyes, but there it was. (I mean, technically, the ROCm stuff is not a driver, although some people may call it that, but that's beside the point--call it what you like; that's not how it works, anymore.)
I don't know how nVIDIA works their drivers, but I have to think that AMD's new stack solves this problem for both teams.
Wait, sorry, I take back my statement that ROCm should solve it. I think what is working so well on Linux may be just the ICD system that allows for multiple OpenCL libs to be installed simultaneously. ROCm is using that mechanism, I think. I checked on my system, and the that old shared lib that used to get overwritten by the driver *does not exist* anymore. But, I think it's the ICD system that allows for that. If Windows doesn't have that, then there's nothing software can do about it; they may still have to overwrite the master OpenCL DLL.
BUT, I would still be disturbed if that is really how it works with the latest Windows, latest hardware, and latest drivers.
It looks like about 30 percent of the top 50 are running Radeon gpus. And about half are running big Navi. The rest are running older gpus. Frequently Radeon VII or variations of them.
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
It looks like about 30 percent of the top 50 are running Radeon gpus. And about half are running big Navi. The rest are running older gpus. Frequently Radeon VII or variations of them.
Let's see... something is wrong with your math.
Of the Top 50 Users in Einstein: 50 X .3 = 15 running Radeon GPUs
Half of Top 50 Users in Einstein: 50 X .5 = 25 running Big Navi GPUs
15 + 25 = 40
The rest of the Top 50 in Einstein: 50 - 40 = 10 running older (Radeon VII or variations of them) GPUs
It looks like about 30 percent of the top 50 are running Radeon gpus. And about half are running big Navi. The rest are running older gpus. Frequently Radeon VII or variations of them.
Let's see... something is wrong with your math.
You did not read it the way he meant it--he meant half of the top 50 AMD units were Big Navi.
It looks like about 30 percent of the top 50 are running Radeon gpus. And about half are running big Navi. The rest are running older gpus. Frequently Radeon VII or variations of them.
...hmm...
Tom M wrote:
I am already fully announced as being brain dead today.
Ahh... blame it on a concussion... who will know?
Tom M wrote:
It won't be the first time something is wrong with my arithmetic reasoning.
And I doubt it will be the last...
archae86 wrote:
GWGeorge007 wrote:
Let's see... something is wrong with your math.
You did not read it the way he meant it--he meant half of the top 50 AMD units were Big Navi.
It may be semantics, but I stand by what I said.
archae86 wrote:
Or was I supposed to detect that you were joking?
...what?... Me... JOKING???
ROFLMAO!!!! Ha Ha Ha LOL!!! Me...??? No way... or maybe... just maybe... a little bit!
Peter van Kalleveen
)
Hmm, are you talking about Windows? That's ironic because I have been fighting with the Linux AMD driver installer for approximately 20 years and, just recently, for my new 6800XT, I discovered that *no drivers* are required to make it work for BOINC. The ancient way of offering OpenCL support/APIs by overwriting libOpenCL.so, as well as the newer way of installing an ICD that allows for multiple OpenCL paths to exist simultaneously--which I assumed was how Windows did it, too--are *gone* because AMD has ended "direct" OpenCL support through their drivers and pushed all OpenCL support on top of ROCm. So, OpenCL doesn't talk directly to the AMD driver anymore. I think it goes kernel/driver/Vulkan/ROCm/OpenCL, now. If you have an RDNA2 card, you shouldn't be installing any OpenCL "drivers". You can even see it in the documentation for ROCm; it describes how the stack has changed with ROCm. The thought that, in Windows, they are still overwriting shared libs, blows my mind. That's a hack from before nVIDIA sold their first card.
Here is mystery I uncovered that proved it: Uninstall all the AMD drivers and install (or leave installed) ROCm, then run whatever program shows you OpenCL supported devices. My card showed up without any divers install. I couldn't believe my eyes, but there it was. (I mean, technically, the ROCm stuff is not a driver, although some people may call it that, but that's beside the point--call it what you like; that's not how it works, anymore.)
I don't know how nVIDIA works their drivers, but I have to think that AMD's new stack solves this problem for both teams.
Wait, sorry, I take back my
)
Wait, sorry, I take back my statement that ROCm should solve it. I think what is working so well on Linux may be just the ICD system that allows for multiple OpenCL libs to be installed simultaneously. ROCm is using that mechanism, I think.
I checked on my system, and the that old shared lib that used to get overwritten by the driver *does not exist* anymore.But, I think it's the ICD system that allows for that. If Windows doesn't have that, then there's nothing software can do about it; they may still have to overwrite the master OpenCL DLL.BUT, I would still be disturbed if that is really how it works with the latest Windows, latest hardware, and latest drivers.
Oh, nope, wrong again; that
)
Oh, nope, wrong again; that file is there. I think it's the entry point to the ICD. Huh, I guess that makes sense.
It looks like about 30
)
It looks like about 30 percent of the top 50 are running Radeon gpus. And about half are running big Navi. The rest are running older gpus. Frequently Radeon VII or variations of them.
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Tom M wrote: It looks like
)
Let's see... something is wrong with your math.
Of the Top 50 Users in Einstein: 50 X .3 = 15 running Radeon GPUs
Half of Top 50 Users in Einstein: 50 X .5 = 25 running Big Navi GPUs
15 + 25 = 40
The rest of the Top 50 in Einstein: 50 - 40 = 10 running older (Radeon VII or variations of them) GPUs
Then just how many are running NVIDIA GPUs?
Proud member of the Old Farts Association
It won't be the first time
)
It won't be the first time something is wrong with my arithmetic reasoning.
I did finger count all of the amd-based gpus in the top 50.
X/50 equals 30 percent.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
GWGeorge007 wrote: Tom M
)
You did not read it the way he meant it--he meant half of the top 50 AMD units were Big Navi.
Or was I supposed to detect that you were joking?
I am already fully announced
)
I am already fully announced as being brain dead today. The only thing I am pretty sure about is there are 15 amd-related gpus in the top 50.
And I think that is an increase.
Any other statement I made may or may not be brain impaired.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Tom M wrote:It looks like
)
...hmm...
Ahh... blame it on a concussion... who will know?
And I doubt it will be the last...
It may be semantics, but I stand by what I said.
...what?... Me... JOKING???
ROFLMAO!!!! Ha Ha Ha LOL!!! Me...??? No way... or maybe... just maybe... a little bit!
Proud member of the Old Farts Association
Now I see the communications
)
Now I see the communications flaw. I meant half of the 30 percent of the top 50 which would mean around 7 Big Navi gpu systems.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)