Thread 'PCI express risers to use multiple GPUs on one motherboard - not detecting card?'

Message boards : GPUs : PCI express risers to use multiple GPUs on one motherboard - not detecting card?
Message board moderation

To post messages, you must log in.

Previous · 1 . . . 6 · 7 · 8 · 9 · 10 · 11 · 12 . . . 20 · Next

AuthorMessage
Ian&Steve C.

Send message
Joined: 24 Dec 19
Posts: 229
United States
Message 95712 - Posted: 5 Feb 2020, 20:33:51 UTC - in response to Message 95710.  

besides the fact that wikipedia is editable by anyone...

again, STANDARDIZED PC COMPONENTS. that's what is being discussed.

a VRM is not a PSU.
ID: 95712 · Report as offensive
Les Bayliss
Help desk expert

Send message
Joined: 25 Nov 05
Posts: 1654
Australia
Message 95714 - Posted: 5 Feb 2020, 21:52:18 UTC

This "discussion" is getting a bit heated. And drifting from at least the last part of the topic title.
ID: 95714 · Report as offensive
Ian&Steve C.

Send message
Joined: 24 Dec 19
Posts: 229
United States
Message 95716 - Posted: 5 Feb 2020, 22:06:50 UTC - in response to Message 95715.  

you could try googling PC VRM and see how many PSUs show up, or vice versa.

there's a reason standardized nomenclature exists.
ID: 95716 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 1301
United Kingdom
Message 95719 - Posted: 6 Feb 2020, 8:00:18 UTC - in response to Message 95707.  

I missed those - some are AC-DC and others are, as you say AC-AC.
ID: 95719 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 1301
United Kingdom
Message 95720 - Posted: 6 Feb 2020, 8:03:35 UTC - in response to Message 95716.  

FUNCTIONALLY a VRM is a power supply, or even a component within a power supply.
ID: 95720 · Report as offensive
ProfileDave
Help desk expert

Send message
Joined: 28 Jun 10
Posts: 2709
United Kingdom
Message 95721 - Posted: 6 Feb 2020, 10:59:31 UTC

Extraneous information aside, I still disagree that a VRM is a PSU. The lack of AC-DC conversion is a major design element that separates the two.


Please don't tell my laptop that the box I connect to the 12V socket in a car isn't a PSU as I would like to keep using it for longer than its internal battery allows when away from mains.
ID: 95721 · Report as offensive
ProfileKeith Myers
Volunteer tester
Help desk expert
Avatar

Send message
Joined: 17 Nov 16
Posts: 890
United States
Message 95723 - Posted: 6 Feb 2020, 17:01:53 UTC - in response to Message 95293.  

but I also suspect you can apply some optimizations (you'd have to inquire at Einstein or MW forums if there any) to make those jobs run faster, at the expense of using more resources.


I shall ask. Although your SETI one is the only one I've ever heard of. What's the name of it so I can refer to it so they know what I'm talking about?

If you are referring to the name of the optimizations for the SETI special application, the one being referenced is -nobs (or no blocking sync)
The mechanism that was being used at GPUGrid for the older deprecated applications was called SWAN_SYNC=1 and did the same thing. No longer used there for the current applications.

What the parameter does is prevent the cpu thread servicing the gpu task from switching away to do other things when listening for the gpu's request for data.

It all depends on how the application developer wrote the science application. Einstein and Milkyway developers wrote their applications differently than what the GPUGrid and SETI developers did.

Different goals and such.
ID: 95723 · Report as offensive
ProfileKeith Myers
Volunteer tester
Help desk expert
Avatar

Send message
Joined: 17 Nov 16
Posts: 890
United States
Message 95724 - Posted: 6 Feb 2020, 20:54:30 UTC - in response to Message 95345.  

Maybe in the past with old tasks Einstein used to have more reliance on PCIe, but it does not appear to be the case anymore.

Yes, the comment about Einstein gpu tasks needing a lot of PCIE bandwidth was in reference to the old deprecated project BRP4G application. That campaign is long ended many years ago. It did suffer a task slowdown if you dropped the card running on it from X16 to X4 for example.

The current MW and Einstein gpu tasks don't seem to work the gpu very hard. Not as hard as Seti gpu tasks for example. I haven't seen any difference in current MW or Einstein gpu applications whether they run at X8 or X4. The task times are always within normal variance.
ID: 95724 · Report as offensive
ProfileKeith Myers
Volunteer tester
Help desk expert
Avatar

Send message
Joined: 17 Nov 16
Posts: 890
United States
Message 95726 - Posted: 6 Feb 2020, 21:26:14 UTC - in response to Message 95376.  

on such a slow system, i think your only shot at running GW would be to run 1 GPU only with 1 GW task and no other CPU tasks running. but maybe not even then.

My one dedicated Einstein system is a 2700X and 3 GTX 1070Ti's. Happily runs Gravity Wave cpu tasks on 8 threads plus a task each on each gpu which hop around between Gravity Wave GPU and Gamma Ray GPU tasks.

No problem supporting the various work on that machine. It is responsive and crunches tasks with no apparent slowdowns. The gpus are running in X8/X8/X4 PCIE bus widths with the X4 slot dropping to Gen. 2 speeds.
ID: 95726 · Report as offensive
Ian&Steve C.

Send message
Joined: 24 Dec 19
Posts: 229
United States
Message 95728 - Posted: 6 Feb 2020, 22:38:21 UTC - in response to Message 95726.  

My one dedicated Einstein system is a 2700X and 3 GTX 1070Ti's. Happily runs Gravity Wave cpu tasks on 8 threads plus a task each on each gpu which hop around between Gravity Wave GPU and Gamma Ray GPU tasks.

No problem supporting the various work on that machine. It is responsive and crunches tasks with no apparent slowdowns. The gpus are running in X8/X8/X4 PCIE bus widths with the X4 slot dropping to Gen. 2 speeds.


well you essentially have 8 modern/fast threads free to service 3 GPUs. which is more than enough. the OP is running a very very old 4c/4t chip. and with the amount of CPU resources needed on that app for nvidia cards, it's not worth running more than 1 task on 1 GPU. the GW app likes to overflow into multiple threads even for just GPU support. i've witnessed the GW tasks on the nvidia app requesting 130-150% of a CPU thread, which causes 2 threads to be used, it doesnt just stop with one thread. and that was on a much stronger CPU still. (E5-2600 v2 Xeons)
ID: 95728 · Report as offensive
ProDigit

Send message
Joined: 8 Nov 19
Posts: 718
United States
Message 95734 - Posted: 7 Feb 2020, 5:32:58 UTC - in response to Message 95671.  

I mean to say, some programs use pcie bus data as a source of GPU utilization. But pcie data transfer can not be measured with any close precision, because such measuring sensors don't exist in common PC hardware. At best, an approximation of 15% can be guessed.

Most GPUs crunching will show 100% utilization, even if they're plugged in a PCIE x1 slot. But programs like folding at home, have an accurate performance sensor, that shows a sharp performance drop on the fastest GPUs on pcie x1 ports, even if the GPU utilization sensors remain at 100%.

More than likely, they're based on GPU frequency, average load, as well as pcie data bus transfer (which some sensors measure as GPU 100% utilization, even if the GPU frequency drops to idle for a moment).


it can be measured with reasonable precision, at least on nvidia gpus. several tools exist that measure this. I've used them (gmonitor on linux). precision is good enough that you see a rough doubling of the PCIe bandwidth use when you halve the link speed or width. its good enough to be able to tell if you're bottlenecked or not, which is all most people care about.

i've never seen any tool read pcie bandwidth and report that value as "gpu utilization".

as far as folding performance, probably best to just read their own forums and what people there have reported: https://foldingforum.org/viewtopic.php?nomobile=1&f=38&t=31708

"My RTX 2080 Ti does seem to be constrained by the 1x bandwidth.
On a 4x/8x/16x slot it gets 2,3M PPD, on a PCIE 3.0 1x slot with this riser, it only gets 1,9-2,1M PPD"

that equates to about 9-17% performance loss, when cutting the bandwidth by over 90%, which means you don't need a huge amount of bandwidth to begin with. I don't think I would call that a sharp drop. seems like the limit is a little over PCIe 3.0 x1. all the way down to x4 seems to be no impact.

but I don't fold anymore (not since like 2005-ish) and don't know if PCIe traffic is used throughout the whole WU, or if it only uses it during certain times.

Boinc is similar to FAH, it will blip a fraction of a second of data, every so often (ten seconds or so) to the GPU via the PCIE bus.
PCIE data can be read by Nvidia drivers, but they're woefully inaccurate. Like mentioned down to ~15% tolerance. That means, if a GPU shows 45% PCIE bandwidth utilization, it's anywhere from 35-55%, which is pretty inaccurate.
Furthermore, the Nvidia drivers are what pass the info on to any third party programs on Linux. They themselves poll the PCIE bandwidth only once every so many seconds, and cut an average.
In the case of Boinc and FAH, this could show 100% and 15% utilization in 2 consecutive readings; or an average reading that's lower than peak PCIE traffic.
Overall though, the Nvidia driver info is good enough to show if the PCIE utilization is higher than 80% (which it is on a PCIE 3.0 x1 slot, feeding a 2080 Ti), you can call it a bottleneck.

FAH Core 21 is close to a lot of Boinc projects (including einstein, the Collatz, Prime and GPU grid....)
Core 22 uses up a lot more PCIE bandwidth (2-4x more). I would suspect that each cluster of data is chopped up more, and the CPU does more copy commands to the GPU, in order to keep all the cores active as much as possible.
But for Boinc, generally:
PCIE 2.0 x1 < GT1030 graphics cards Windows
2.0 x4 , or 3.0 x1 <GTX 1060 graphics cards in Windows
3.0 x1 <RTX 2060 Super (on Linux)
3.0 x4 <RTX 2080 Ti on Linux or <2060 Super on Windows
3.0 x8 <RTX 2080 Ti on Windows

I wanted to buy another 2080 Ti, but I guess I'll just wait for the 3000 series GPUs to come out, and hope they won't be much more expensive than the 2080 Ti.
ID: 95734 · Report as offensive
ProfileJoseph Stateson
Volunteer tester
Avatar

Send message
Joined: 27 Jun 08
Posts: 641
United States
Message 95736 - Posted: 7 Feb 2020, 5:59:56 UTC - in response to Message 95729.  


I can't even service one GPU with that CPU. It's doing MW and Einstein Gamma instead. No matter how many Gravity tasks I put onto one GPU, the CPU can't keep up. I can max out all four cores of the CPU, while the GPU sits at 30%


You mention gravity and also gamma. I quit doing gravity due to low credit and problems with drivers on older boards. With "weak" CPU, the gamma ray pulsar seems to run nicely on ATI boards, not so on NVidia. This was discussed over at Einstein and it seems to be the way the polling is implemented as well as the hardware.

Following is a Linux system with a really cheap 2 core Celeron G1840. Motherboard is Hl61BTC I got down from the attic and put a pair of RX560 and a pair of RX570

The 11 minute 44 sec completion is typical for the RX570, not obvious from the picture, but the RX560 typical is 25 - 29 minutes. Note the CPU usage is at most %22 which works out nicely for the celeron.



On the other hand, the nvidia board uses up a full cpu. Both systems run 18.04

ID: 95736 · Report as offensive
Ian&Steve C.

Send message
Joined: 24 Dec 19
Posts: 229
United States
Message 95744 - Posted: 7 Feb 2020, 16:51:07 UTC - in response to Message 95734.  

Boinc is similar to FAH, it will blip a fraction of a second of data, every so often (ten seconds or so) to the GPU via the PCIE bus.
PCIE data can be read by Nvidia drivers, but they're woefully inaccurate. Like mentioned down to ~15% tolerance. That means, if a GPU shows 45% PCIE bandwidth utilization, it's anywhere from 35-55%, which is pretty inaccurate.


"BOINC" does nothing of the sort. BOINC doesn't use the GPU, the science application does. and how it utilizes resources is entirely up to how that application was written. The Einstein GPU apps use system resources entirely different than the CUDA special application at SETI as an example, but both run on the BOINC platform.

I've measured PCIe resources being used on both of these apps, and SETI CUDA special app so far is the only one with a constant stream of data being sent over the PCIe bus for the entire duration of the WU run. using roughly 600MB/s of the bandwidth all. the. time. Einstein almost uses none, only like 1% used.

can you provide a source from nvidia that claims PCIe bandwidth is only accurate to 15%? or is that number purely speculative? I have not found anything like that in their documentation.
ID: 95744 · Report as offensive
Previous · 1 . . . 6 · 7 · 8 · 9 · 10 · 11 · 12 . . . 20 · Next

Message boards : GPUs : PCI express risers to use multiple GPUs on one motherboard - not detecting card?

Copyright © 2024 University of California.
Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.