Running CPU tasks on a GPU project is a big no, no. Even the fastest server CPU would get absolutely demolished by any AMD/Nvidia GPU of recent years. Your RAC would stay at what it is with or without running CPU tasks.
On a side note, I'm wondering, is it possible to crunch on your behalf from my account, by assigning your CPID as an external one?
I have GPUs to spare, so if you post your CPID here, I would give you a free boost.
I have dropped the "run GPU tasks on the CPU" setting for that profile.
I know "we" have crunched on another's behalf near the end of Seti@Home using that trick. I am hazy on the details but confident it could be done.
The system that is the center of this discussion is my Daily Driver running under Windows 11. I thank you for the offer. I am wondering if my Daily Driver would benefit from a boost or just confuse me :)
Now if you wanted to help me give Ian&SteveC a run for his money for first place I can see all sorts of "glory" in that :) But probably need to confirm I can get back up to above 6M+ RAC on that machine first :) And confirm that it works that way. I know we used that trick to raise his total production. Don't remember if it effected his RAC or not.
I am wondering if I allow only 2 "Multi-Directional Gravitational Wave search on O3 (CPU)" to run at a time if they will run 1/3 slower rather than more than twice as slow. And still not slow the other e@h CPU tasks down. Plus the monkey wrench in the experiment by replacing the cpu/iGpu (5700G).
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 am wondering if I allow only 2 "Multi-Directional Gravitational Wave search on O3 (CPU)" to run at a time if they will run 1/3 slower rather than more than twice as slow. And still not slow the other e@h CPU tasks down. Plus the monkey wrench in the experiment by replacing the cpu/iGpu (5700G).
Since WCG has been down, I crunched CPU tasks here for a little bit to see how they run on the new Threadrippers. I can tell you that I observed some "interesting" results when I had the 64 core/128 thread TR running ~120 of these tasks concurrently. The more you run, the more they slow down (and I am not referring to the CPU speed, overall- it still reported the cores were running at ~3.4GHz). And the longer I let them run, the worse it became. I could stop half of them, and it still would not speed up the other tasks. It was like an exponential decrease over time- they would all start fast, then just continue to slow down until they basically didn't move. It was interesting to watch the 64 core system bottleneck. There was still memory available but it used over half of what was available.
Anandtech wrote an article, years ago (this one https://www.anandtech.com/show/15483/amd-threadripper-3990x-review/2) describing the issue you are experiencing. The more cores you have at full load the lower the power consumption per core is. And obviously, the performance goes down, as there is a limited TDP for the entire package.
I am wondering if I allow only 2 "Multi-Directional Gravitational Wave search on O3 (CPU)" to run at a time if they will run 1/3 slower rather than more than twice as slow. And still not slow the other e@h CPU tasks down. Plus the monkey wrench in the experiment by replacing the cpu/iGpu (5700G).
Got that started. Only two MDGWO3 are now running on the Windows box.
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 believe I also had a similar issue under a Ryzen 3950x with Rosette at Home way back when. I could run no more than 4-6 cpu tasks without the all the cpu processing slowing way down.
Haven't been happy with running those tasks since they switched to a VM-base cpu processing. Something about my rigs became unstable.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Has anyone done a Tasks Retired test while adding FRGP6 WUs one-by-one?
This proved useful in determining number of threads to use on multi-threaded Milkyway tasks. I believe it's related to overloading the L3 cache. I've noticed that RAM is not the limiting factor when running multiple FRGP5 WUs. CPU utilization slows down long before RAM is filled.
Dido wrote: @TOM M Running
)
I have dropped the "run GPU tasks on the CPU" setting for that profile.
I know "we" have crunched on another's behalf near the end of Seti@Home using that trick. I am hazy on the details but confident it could be done.
The system that is the center of this discussion is my Daily Driver running under Windows 11. I thank you for the offer. I am wondering if my Daily Driver would benefit from a boost or just confuse me :)
Now if you wanted to help me give Ian&SteveC a run for his money for first place I can see all sorts of "glory" in that :) But probably need to confirm I can get back up to above 6M+ RAC on that machine first :) And confirm that it works that way. I know we used that trick to raise his total production. Don't remember if it effected his RAC or not.
I am wondering if I allow only 2 "Multi-Directional Gravitational Wave search on O3 (CPU)" to run at a time if they will run 1/3 slower rather than more than twice as slow. And still not slow the other e@h CPU tasks down. Plus the monkey wrench in the experiment by replacing the cpu/iGpu (5700G).
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'm certainly happy to
)
I'm certainly happy to give Ian&SteveC a boost to first place here at einstein@home. I don't know though, if he would accept any help from me.
Haha. Appreciate the
)
Haha. Appreciate the offer, but I’m content with second place under my own power :)
_________________________________________________________________________
Dido wrote: I'm certainly
)
Ah, different ranking. I was talking about "top 50" and you were talking about top users.
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: I am wondering
)
Since WCG has been down, I crunched CPU tasks here for a little bit to see how they run on the new Threadrippers. I can tell you that I observed some "interesting" results when I had the 64 core/128 thread TR running ~120 of these tasks concurrently. The more you run, the more they slow down (and I am not referring to the CPU speed, overall- it still reported the cores were running at ~3.4GHz). And the longer I let them run, the worse it became. I could stop half of them, and it still would not speed up the other tasks. It was like an exponential decrease over time- they would all start fast, then just continue to slow down until they basically didn't move. It was interesting to watch the 64 core system bottleneck. There was still memory available but it used over half of what was available.
Anandtech wrote an article,
)
Anandtech wrote an article, years ago (this one https://www.anandtech.com/show/15483/amd-threadripper-3990x-review/2) describing the issue you are experiencing. The more cores you have at full load the lower the power consumption per core is. And obviously, the performance goes down, as there is a limited TDP for the entire package.
Tom M wrote: I am wondering
)
Got that started. Only two MDGWO3 are now running on the Windows box.
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 believe I also had a
)
I believe I also had a similar issue under a Ryzen 3950x with Rosette at Home way back when. I could run no more than 4-6 cpu tasks without the all the cpu processing slowing way down.
Haven't been happy with running those tasks since they switched to a VM-base cpu processing. Something about my rigs became unstable.
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:I have dropped
)
I can't find this in my Project Preferences. Do you mean "Run CPU versions of applications for which GPU versions are available"?
Has anyone done a Tasks
)
Has anyone done a Tasks Retired test while adding FRGP6 WUs one-by-one?
This proved useful in determining number of threads to use on multi-threaded Milkyway tasks. I believe it's related to overloading the L3 cache. I've noticed that RAM is not the limiting factor when running multiple FRGP5 WUs. CPU utilization slows down long before RAM is filled.
https://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=5028&postid=76264#76264