Feeding a 1070 Ti with a tiny processor

[Paragon]

I'm noticing something interesting on my new Linux super low-power build (25 watts idle without GFX card installed). When doing GPU folding with my Asus GeForce GTX 1070 Ti, I am only getting 200-500k PPD depending on WU. On my Windows 10 machine (AMD FX-8320e, 16 GB Ram) this card did around 700K PPD average. I had expected Linux to be faster. Is it possible this system is too slow to feed the card?

System specs:

AMD Athlon 5350 APU (4 cores, 2.05 GHz, 25 watts TDP)

Gigabyte GA-AM1M-S2H microATX board

8 GB DDR3l (1.35 volt) memory

Asus GeForce 1070 TI (running proprietary PPA 410 long term support drivers)

Seasonic Focus 80+ Platinum 550 watt PSU

Linux Ubuntu 18.04 (also tried Mint 19.1 and Lubuntu 18.10)

Things I've tried to speed it up:

-Upgraded proprietary Nvidia drivers after purging open source nouveau drivers
-Installed latest OpenCL package
-Flashed bios to latest version (F2)
-Tried Mint 19.1 instead (same issue)
-Set power profile to performance in Nvidia x-settings
-Disabled cool and quiet / CPU low power state in the BIOS

I'm using FAH Client and Control version 7.5.1. Watching the GPU load in Nvidia settings shows variation between 70-100 percent with occasional dropouts to 20 percent. GPU is very warm but not hot (78 degrees c, fan at 60 percent). Pci-e shows 8 percent bus utilization. GPU core boosted itself to around 1810 MHz. CPU shows 1 core pegged at 100 percent use, the rest 1-10 percent.

Any thoughts?

[bruce]

Are you using the same UserName and passkey on all the systems you mentioned?

[foldy]

The problem is with your mainboard only supporting pcie x16 slot with pcie 2.0 x4 speed which bottlenecks on Windows. So you need to run Linux for full speed. I see you already tried Linux Mint and had the same issue but I guess it is a different issue there which maybe can be fixed. The CPU type is very slow, half the speed of Celeron G1820 even with 2 Ghz and maybe it is possible to overclock the CPU to see if it makes a difference? But I guess this CPU would need to run at 4 Ghz to not bottleneck FAH. FAH only uses single core performance to feed the GPU.

[Joe_H]

The CPU is used for two separate functions during GPU folding. The first has already been mentioned, creating the packets of data to be processed by the GPU and retrieving them.

The second is to create periodic checkpoints and run sanity checks on the WU data that has been processed to that point. Your 5 year old low end 5350 APU will spend a significantly longer time doing those compared to a more recent or faster processor.

[Paragon]

Yeah I thought about over clocking it to see if that reveals the bottleneck but this board has no OC options. I also have a newer A10 Apu (quad core, 3.8 GHZ, 95w TDP) on a different board. I guess I'll try the exact same install on that hardware and see if it works better.

I'm just a bit surprised because I've seen people on here say you can do GPU folding with the cpu equivalence of a potato...maybe I've found the limit?

[foldy]

For the current fast GPU generation which I count gtx 1070 Ti to, yes you found the limit. A10 3.8 GHZ looks better for CPU speed (but dual core would be enough for one GPU). But still mainboard may deliver pcie 2.0 x4 speed only - so you need Linux again.

[Paragon]

Update:

I might not be CPU bottlenecked after all. Well, I am, but it's a software issue. For some reason, it is running at 800 MHZ instead of 2.05 GHz when Ubuntu is loaded (BIOS shows 2.05 GHz at boot). I used the cpufreq-info command to see the clocks in the OS. Apparently the "ondemand" governor doesn't clock the CPU up at all while GPU folding. Some gamers had a similar problem here

https://forum.level1techs.com/t/the-ubu ... fix/129839

I tried that fix plus a bunch of others but nothing works...CPU stays stuck at 800 MHz. Not enough to feed the 1070 TI...

Anyone else have this happen? What can I do? I'm guessing if I can get Ubuntu to run the processor at 2.05 GHz, the GPU PPD should go up (basically I'm feeding it with 1 800 MHz core and that's not enough for a 1070 Ti)

[bruce]

It's not a problem I've ever encountered so any suggestions I might offer would be a WAG.

What happens if you give the CPU some work to do? ... like maybe allocating a CPU slot using maybe half of your threads (depending on how many your CPU supports, of course. Apparently the governor doesn't perceive that the CPU has enough work to so that 800 MHz can keep up with what work it sees. (Of course the performance patch discussed in that topic probably works at least as well.)

[antropofob]

I am using https://launchpad.net/indicator-cpufreq to downclock my CPU so maybe you could use it to do the opposite.

[foldy]

http://www.thinkwiki.org/wiki/How_to_use_cpufrequtils

[MeeLee]

A 2Ghz CPU usually is good enough to feed a 2070.
Of each core, you'll probably see 100% usage, per card; however, I think a 2Ghz processor will use more like 30-40% of actual work. The rest is some sort of 'CPU thread lock' or something, that Nvidia drivers try on the CPU, to keep one GPU assigned to one thread.

800Mhz should significantly lower your PPD though...

[bruce]

THe fundamental issue is that neither you nor your OS know how much of a CPU is needed -- and the task priority in the OS can't easily be used.

MeeLee's characterization of 2 GHz vs. 800 MHz, but it could easily be either right or wrong -- we just don't know. Then, too, it's going to be different for Windows vs. Linux vs. MacOS-X.

The driver uses what's called a "spin wait" which artificially keeps one CPU thread busy, whether it's essential to the GPU's performance or not so the OS reports that it uses 100% of a single thread.

ASSUMING those numbers are correct, then I can guess that there are cases where 1.4 GHz is essential. If you have a 0.8 GHz CPU thread, then your GPU is going to be waiting about for the CPU to supply data about 30% of the time. If you have a 2GHz, the CPU will be doing essential work about 70% of the time and will waste the other 30%. Both will appear to be 100% busy.

In the 2.0 GHz case, there might be obscure ways to have your CPU release the allocate the excess 30% to be used for something else (depending on your OS ) but if you try to do that, the resource competition is probably going to slow down your GPU.

[Theodore]

Taskmanager in Windows, and system monitor in Ubuntu and Lubuntu, or htop in Linux terminal, will show you how much of the CPU is dedicated to locking the CPU thread, and how much (kernel time) is used to actual data.

In htop it shows as the primary color of CPU utilization, and the thread locking (idle data) shows up as the secondary (you can assign any color you like)



The above imagine is purely for illustration purposes, and not an actual image of CPU usage under Fah.

You'll find, that lowering CPU frequency, will shorten the secondary color (in this case green), and increase the primary color (in this case blue).

Lowering CPU speed to the point where the kernel times reach almost 100%, does very little to PPDs.
Lowering them further, and PPDs drop drastically.
It's a good way to see how much CPU speed your GPU really needs, and a potential way to save power, by underclocking your CPU, if the motherboard allows it.


If you see a situation like above either thread 5, 8, 10, 11, 15, or 16, your GPU is being limited by the CPU. This can be either your CPU is too slow, or other processes fully utilize the CPU, which potentially can slow down the GPU.

In most cases, you'll find your graphics cards utilizing a similar percent of CPU as thread 2, 6, 7, 13, or 14.

The 'idle data' sent by the CPU can be utilized by other software (like boinc) as well; and often results in only a minor performance penalty on the GPU.

[Paragon]

So I double checked a bunch of settings...user name and passkey all check out. System monitor shows the CPU core feeding the card @ 800 MHz is taking an absolute beating. I tried CPU folding (4 cores at full) and it didn't make the CPU speed increase. Interesting side note was that system power consumption when CPU folding (GPU paused) only increased by 7 watts at the wall vs CPU idle. Even for a 20 watt TDP APU, this seems small...further evidence that the CPU is not clocking up.

I used cpufreq-info and cpupower commands to successfully change the governor mode from "ondemand" to "performance", but unlike many of the threads online, in my case it didn't fix anything. Even in performance mode it stays at 800 MHz.

I think I'm going to try a similar Linux install on a different machine (A10 Apu, Asrock board) and see if this is somehow hardware dependent. The Athlon 5350 / AM1 socket is pretty obscure. Maybe there is some strange incompatibility happening.

[foldy]

Did you try to force CPU clock?

Code: Select all

sudo cpufreq-set -g userspace
sudo cpufreq-set -f 2000Mhz