A minor correction to my previous post. While the command line I posted worked, what actually happened is that, because the --opencl=pal option is no longer supported, the installer replaced it with --opencl=rocr. The AMD documentation alludes to this in its installation command example of:
./amdgpu-pro-install -y --opencl=rocr,legacy --headless
Note that the --no-dkms option that I posted doesn't work with the amdgpu-pro-install command, however, only with amdgpu-install.
By the way Keith, thanks for that tip on dkms uninstall.
The point of all that being, while it may be a coincidence, it's looking like the new (for me) ROCr implementation for OpenCL has given my RX 5600 XT a big boost in performance for crunching gamma-ray tasks. With the older amdgpu driver version 20.20 (Radeon™ Software for Linux® Driver for Ubuntu 20.04.1), my average individual task time for 3X tasks was about 6 minutes (reported as 1080 s avg. for three concurrent tasks). After today's driver upgrade to version 20.45, task time is averaging about 4:20 (reported as ~780 s for three concurrent tasks). I can't at the moment think of what other variables might be related to that awesome performance increase.
Ideas are not fixed, nor should they be; we live in model-dependent reality.
That is what should be expected with the ROCr implementation of OpenCL.
Much more modern and efficient implementation compared to the legacy OpenCL PAL component.
While I have no personal experience with AMD gpu hardware, I do follow all the OpenCL AMD development articles at Phoronix. AMD does strongly believe in open source development. Much more so compared to Nvidia.
You might find the latest RX6000 OpenCL tests an interesting read.
... while it may be a coincidence, it's looking like the new (for me) ROCr implementation for OpenCL has given my RX 5600 XT a big boost in performance for crunching gamma-ray tasks....
Alas, that party's over. My average individual task times are back to a tad under 6 min, so only a very minor improvement with the ROCr drivers. We'll see if the performance variability continues. Yesterday's short times included a mix of task IDs, as did the previous and current "long" tasks, so nothing obvious (to me) to explain what happened. It was fun while it lasted.
Ideas are not fixed, nor should they be; we live in model-dependent reality.
It looks like I have a stable Windows 10 system with two Rx 5700's now on my Ryzen 3950x box. The RAC is on a steady upward slope.
Would a user of Windows 10 and Rx 5000 series talk about the utilities and tuning they are using to maximize Rx 5000 series production, please?
Right now I am single-tasking the GPUs because it is not even close to the guesstimated peak. But I am going to want to "squeeze the last drop" out of this rig.
I understand that I will (eventually) be running ~3 gpu tasks per gpu card (maybe 2 and probably not 4).
There are the Windows Radeon tools with "automatic" overclocking of the GPU, GPU memory, and undervolting the GPU.
So far my experiments using the "automatic" settings have resulted in instability of the video cards.
For a non-Linux system can you change the state-level of a Radeon gpu card? I think I read someone was getting more production at one step below the maximum state? What utility?
Anyway.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
My google news feed (on my cellphone) predicted that we would have "another" runup in gpu prices for the older generation of Rx 5000 series due to bitcoin mining pressure after the newest generation of AMD GPUs started shipping.
I was hoping for a price collapse on the Rx 5700-non XT cards instead.
Currently, it appears the Rx 5700 xt's are cheaper than the 5700's which is just a bit counter-intuitive :)
Is anyone seeing a source of Rx 5700's that are below $350???
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
On another thread and on another project Keith Myers proposed that for some gpu applications overclocking the VRAM would result in faster processing vs. overclocking the gpu.
Do we have any evidence on this at E@H?
I have already demonstrated that the automatic OC of VRAM for R5700's results in computation errors. So I am going to have to try some reduced manual OC of VRAM.
The Automatic OC of the gpu has not (yet) resulted in any computation errors.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
I think Ian posted once that a memory OC on Einstein has minimal effect. Not sure if I remember that correctly. Would have to search posts.
Memory errors are usually because the memory is getting too hot. So if the card temp is iffy at best already, a very mild memory OC should be used.
On my GTX 1080 Ti hybrid cards, I have a 1400 Mhz memory OC to get back to 11400Mhz effective clock rate. About 400Mhz past the gaming P0 clock rate. But they run in the mid to high 40's for temps too.
I don't generate any task errors. I have pushed to 2000Mhz OC before without errors also when I was running Seti.
This morning I was successful in getting 5 tasks to run on my XFX Rx 5700 that is currently running as a single gpu on this system.
The RAC on this system is floating along and slowly decreasing at just under 1,000,000 RAC. It is even probable that it is going to fall out of the top 50 listings.
So I am "playing" with it to see how much total production I can squeeze out of it. The baseline is 3 tasks. It runs 4 tasks but based on another user of Rx 5700's report from 3 tasks to 4 tasks only gains a few %.
So I said, why not check out 5 like I was trying out on a Radeon VII at one time?
The trick btw on both the Rx 5700 and the Radeon VII is not to set the gpu usage to 0.20 but to 0.19. At 0.20 it stays are 4 tasks for some reason.
"Still flogging the willing horse after all these years...."
Depending on a multi-day test run I will step it up or down a task to see if I can get a higher RAC. Also depending on if it crashes.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
You can remove unused dkms
)
You can remove unused dkms driver modules to get rid of that issue.
dkms uninstall -k {driver name} -v {kernel version}
dkms uninstall -m {module name} -v {kernel version}
A minor correction to my
)
A minor correction to my previous post. While the command line I posted worked, what actually happened is that, because the --opencl=pal option is no longer supported, the installer replaced it with --opencl=rocr. The AMD documentation alludes to this in its installation command example of:
./amdgpu-pro-install -y --opencl=rocr,legacy --headless
Note that the --no-dkms option that I posted doesn't work with the amdgpu-pro-install command, however, only with amdgpu-install.
By the way Keith, thanks for that tip on dkms uninstall.
The point of all that being, while it may be a coincidence, it's looking like the new (for me) ROCr implementation for OpenCL has given my RX 5600 XT a big boost in performance for crunching gamma-ray tasks. With the older amdgpu driver version 20.20 (Radeon™ Software for Linux® Driver for Ubuntu 20.04.1), my average individual task time for 3X tasks was about 6 minutes (reported as 1080 s avg. for three concurrent tasks). After today's driver upgrade to version 20.45, task time is averaging about 4:20 (reported as ~780 s for three concurrent tasks). I can't at the moment think of what other variables might be related to that awesome performance increase.
Ideas are not fixed, nor should they be; we live in model-dependent reality.
That is what should be
)
That is what should be expected with the ROCr implementation of OpenCL.
Much more modern and efficient implementation compared to the legacy OpenCL PAL component.
While I have no personal experience with AMD gpu hardware, I do follow all the OpenCL AMD development articles at Phoronix. AMD does strongly believe in open source development. Much more so compared to Nvidia.
You might find the latest RX6000 OpenCL tests an interesting read.
Radeon RX 6800 Series Has Excellent ROCm-Based OpenCL Performance On Linux
cecht wrote:... while it may
)
Alas, that party's over. My average individual task times are back to a tad under 6 min, so only a very minor improvement with the ROCr drivers. We'll see if the performance variability continues. Yesterday's short times included a mix of task IDs, as did the previous and current "long" tasks, so nothing obvious (to me) to explain what happened. It was fun while it lasted.
Ideas are not fixed, nor should they be; we live in model-dependent reality.
It looks like I have a stable
)
It looks like I have a stable Windows 10 system with two Rx 5700's now on my Ryzen 3950x box. The RAC is on a steady upward slope.
Would a user of Windows 10 and Rx 5000 series talk about the utilities and tuning they are using to maximize Rx 5000 series production, please?
Right now I am single-tasking the GPUs because it is not even close to the guesstimated peak. But I am going to want to "squeeze the last drop" out of this rig.
I understand that I will (eventually) be running ~3 gpu tasks per gpu card (maybe 2 and probably not 4).
There are the Windows Radeon tools with "automatic" overclocking of the GPU, GPU memory, and undervolting the GPU.
So far my experiments using the "automatic" settings have resulted in instability of the video cards.
For a non-Linux system can you change the state-level of a Radeon gpu card? I think I read someone was getting more production at one step below the maximum state? What utility?
Anyway.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
Oh yes. Merry Christmas ;)
)
Oh yes. Merry Christmas ;)
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
My google news feed (on my
)
My google news feed (on my cellphone) predicted that we would have "another" runup in gpu prices for the older generation of Rx 5000 series due to bitcoin mining pressure after the newest generation of AMD GPUs started shipping.
I was hoping for a price collapse on the Rx 5700-non XT cards instead.
Currently, it appears the Rx 5700 xt's are cheaper than the 5700's which is just a bit counter-intuitive :)
Is anyone seeing a source of Rx 5700's that are below $350???
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
On another thread and on
)
On another thread and on another project Keith Myers proposed that for some gpu applications overclocking the VRAM would result in faster processing vs. overclocking the gpu.
Do we have any evidence on this at E@H?
I have already demonstrated that the automatic OC of VRAM for R5700's results in computation errors. So I am going to have to try some reduced manual OC of VRAM.
The Automatic OC of the gpu has not (yet) resulted in any computation errors.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
I think Ian posted once that
)
I think Ian posted once that a memory OC on Einstein has minimal effect. Not sure if I remember that correctly. Would have to search posts.
Memory errors are usually because the memory is getting too hot. So if the card temp is iffy at best already, a very mild memory OC should be used.
On my GTX 1080 Ti hybrid cards, I have a 1400 Mhz memory OC to get back to 11400Mhz effective clock rate. About 400Mhz past the gaming P0 clock rate. But they run in the mid to high 40's for temps too.
I don't generate any task errors. I have pushed to 2000Mhz OC before without errors also when I was running Seti.
This morning I was successful
)
This morning I was successful in getting 5 tasks to run on my XFX Rx 5700 that is currently running as a single gpu on this system.
The RAC on this system is floating along and slowly decreasing at just under 1,000,000 RAC. It is even probable that it is going to fall out of the top 50 listings.
So I am "playing" with it to see how much total production I can squeeze out of it. The baseline is 3 tasks. It runs 4 tasks but based on another user of Rx 5700's report from 3 tasks to 4 tasks only gains a few %.
So I said, why not check out 5 like I was trying out on a Radeon VII at one time?
The trick btw on both the Rx 5700 and the Radeon VII is not to set the gpu usage to 0.20 but to 0.19. At 0.20 it stays are 4 tasks for some reason.
"Still flogging the willing horse after all these years...."
Depending on a multi-day test run I will step it up or down a task to see if I can get a higher RAC. Also depending on if it crashes.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!