Maximizing AMD/ATI gpu tasks and Nvidia gpu tasks on the same system
30 Apr 2020 2:24:26 UTC
Topic 222255
(moderation:
I now have a mixed Amd/Ati and Nvidia rig:https://einsteinathome.org/host/12828668
I am wondering 2 related questions.
With the Rx 560 video cards, I have what is the best combination of # of tasks per gpu? And # of tasks per cpu?
I have switched back to a 1:1 ratio for the time being.
The 2nd question is since anything I do to the AMD/Ati cards will also replicate on my Nvidia GPUs, what is the likely impact? Neutral, Negative, or Positive on the Nvidia cards?
I have a choice. I can move all my Nvidia cards to another Nvidia system. And crunch with just the 6 Rx 560 cards on this system.
What will maximize my total production (ignoring any CPU processing of either type of task)?
Thank you.
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!
Interesting results. I
)
Interesting results. I thought the AMD cards were supposed to be much faster on Einstein work like at Milkyway.
Right now your RX560 are taking 1500 seconds per GW task and your 1660 is taking 900 seconds per GW task.
I would guess by increasing to 2X per card you would get more production out of the host since the doubling of time on the Nvidia cards would not be so much more than what the AMD cards are doing.
[Edit] I probably would not go any further because of the limited memory on the RX560 cards which I think is only 4GB. The Velajr1 tasks we have been getting have needed as much as 1.7GB of memory to run. So 2 up would be the safest amount to run.
you can double or triple up
)
you can double or triple up the tasks on AMD cards since the AMD apps here don't use nearly as much CPU as the nvidia cards do. trying to do that on nvidia will result in much more CPU use and just isnt worth it.
Tom, you're trying to do too much at once. It's not feasible to have one system running several different projects with many different cards types/vendors, AND still expect it to be well optimized. You should just simplify/homogenize your systems and you will run things much more smoothly.
make an AMD-only system and run it on the projects that benefit from AMD cards.
make an Nvidia-only system and run it on the projects that benefit from Nvidia cards.
make a CPU-only system that does not use the GPU, and run it on CPU-only projects.
either physically seperate systems, or logically separate by running custom <ignore...> flags in your cc_config(s) and run multiple clients to keep them isolated.
only then will you be able to have things fully optimized without conflicts.
_________________________________________________________________________
Tom M wrote:With the Rx 560
)
You can solve the problem of running different numbers of tasks on different platforms by switching to an app_info.xml. That will allow you to run 2 on the AMD cards and one on the NV, which would be the best numbers.
As for your choice of AMD GPUs, I'd suggest using at least RX570 as they are about twice as fast as a 560 for just a little more money. A RX560 is similar to a NV GTX1060, the 570 is similar to a GTX2070 for much less money.
I gave running the platforms together again on my ASUS Z170-WS with the same results as before, as soon as you add the NV GPU to a system with the AMD driver installed, Ubuntu refuses to boot. I suspect I'll have to try it on the MSI Mining board to have it work. I'll have to round up another power cable adapter for that, so it might be a little while yet.
TBar wrote:You can solve
)
you mean app_config.xml? He shouldn't need to switch to Anonymous Platform just to do this. the documentation for app_config/ Project level configuration shows the ability to control by app_name combined with plan_class, not just short app name, which would control AMD/Nvidia exclusively since the plan_class will be different. use <app_version> instead of <app>.
https://boinc.berkeley.edu/wiki/Client_configuration
_________________________________________________________________________
Thank you all for weighing
)
Thank you all for weighing in.
First approximation of changes will be to consolidate all Nvidia's onto the 3950x box. I am getting good results running at 1 cpu to 0.33 gpu so having enough threads on a 8c/16t cpu would be a problem. I will throttle back the Rosetta@Home to free up enough threads for the 5 gpus (2 RTX 2060 Supers, 3 Gtx 1660 Supers).
I will run 3 Pulsar#1 tasks and 2 Gravity Waves per Rx 560 on the i9 9900k. I found another cheap source of Rx 560's so I will end up with 9 when the next batch arrives.
The i9 will have enough threads left over to run a couple of single thread projects and some R@H which will give more close to the 17+ thread count I have been running only across two boxes.
I will consider running a pure cpu box but I was/am trying to limit myself to two crunchers maximum. Even though I now have a perfectly good Amd 2700 and a spare low end gpu (not good for crunching).
per Rx 560Decisions, decisions.
I want to thank all of you for helping me sort out what I should do.
It does sound like I should "migrate" to Rx 570's which are running about $100 used. I will look at that after I get the rest of the Rx 560's in. Maybe I can trade up on a 2 for 1 ratio with a little cash. That would give me 4 Rx 570's.
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 have an Nvidia only system
)
I have an Nvidia only system and an AMD only system. After farting around (see another tagline about old farts) I have set both systems to 1:1 for the CPU/gpu ratio.
I have also switched both the Gravity Wave gpu only because right now it looks like my system(s) are processing them faster than the Pulsar#1 gpu tasks.
Does anyone have an opinion "the best" ratio for Nvidia CPU/gpu ratios for either (Pulsar#1/Gravity Wave) task? Most gpu production for least CPU thread count?
Yes, I understand it is "hardware-specific" I am just trying to get something more than 0.5 gpu + 0.1 CPU which I suspect is about as good as it is going to get unless you go with 1 gpu + 0.1 CPU.
What are you running?
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 keep forgetting you can't
)
You keep forgetting you can't compare apples to apples, when comparing Nvidia apps to AMD apps.
AMD apps can get away with using a fraction of a cpu core when the Nvidia app is going to ALWAYS require a full cpu core to support each task.
And your comment about GR gpu apps being slower than the GW apps is confusing. Depends on whether you are speaking of the Nvidia card host or the AMD card host.
The GR gpu tasks are faster on the Nvidia cards and slower on the AMD cards. Can't compare a 1660 Super against a RX560. In different leagues.
I seem to be making a little
)
I seem to be making a little progress. Apparently you need to use the latest AMD driver 20.10 for 18.04. Unfortunately it appears you have the same trouble as with the older FGLRX driver, NVIDIA Settings doesn't work when the AMD driver is installed. Next I'll see if there is some way to avoid that major problem. Before that, I had tried getting the older driver to work on the MSI board first, and kept getting some BIOS Error about some "Windows Features" that I couldn't sort out. A BIOS Error, from just installing a Standard RX570? I'm very unimpressed with the MSI BIOS. Since that seemed to be going nowhere, I decided to just add another GTX 1070 to the MSI board and play around with the AMD card on the ASUS board.
For now I have the ASUS-WS board with the RX570 and a nVidia 1050 GPU showing up in BOINC, but No NV Settings working... IF I can get NV Settings to work, I'll try installing the AMD Card in the Machine where I stole the 1070.
Keith Myers wrote: You keep
)
I am now running 1 thread per Nvidia gpu task. One of my questions is how many gpu tasks can I run on a Nvidia? Is anyone running say two tasks on their Nvidia and gaining any production?
I am testing running 0.1 cpu on the Amd cards since they are always listed as "stopped" on the task manager. But can I get more total production with 1, 2 or 3 tasks?
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!
TBar wrote:I seem to be
)
While I did manage to get Nvidia and AMD cards to share the AMD OCL because of the different performance characteristics I was advised to consolidate my Nvidia and AMD cards onto separate machines.
I have done so and am MUCH happier with trying to juggle the parameters with both cards.
If you start with a clean Linux OS install apparently if you install the AMD drivers with OCL first and then install the Nvidia drivers without installing the separate Nvidia OCL driver for Nvidia you may be able to get them to process OCL together.
Unfortunately, my experience was with the older version (18?) that was recommended by the AMD website for an Rx 560 gpu.
Respectfully,
Tom Miller
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!