Are these Power limits (Ubuntu,RTX 3080,3070 Ti's) Correct?

[dfonda]

Finally going to try power Power Limits on the 2 Linux PC's. Gotta get the heat down...I have this from HayesK. Are they still correct amounts to fully perform on my 3080's and 3070's?

Originally from this post in this thread...

Question on upgrade?

Anyone think I would see a noticeable increase in PPD if I upgraded the 3rd PC in my sig to a new system? I was thinking along the lines of a Ryzen 5 5600X, X570, 32 (Or 16) GB of RAM....and a cheap cooler. Keeping the case , 1300W PSU , SSD, either a free Win 10 or Linux. I think that covers...

www.overclockers.com

GPU power limit adjustment

# enable limit adjustment
sudo nvidia-smi --persistence-mode=1

#set GPU 0 power limit 150
sudo nvidia-smi -i 0 --power-limit=150

#set GPU 1 power limit 160
sudo nvidia-smi -i 1 --power-limit=150

# view power limits
nvidia-smi

 

[HayesK]

Yes, that will work for GPU 0, executed in a terminal window, then the terminal can be closed.

For a dual gpu host, I have combined the commands into a single line using &&, which looks like the following:

sudo nvidia-smi --persistence-mode=1 && sudo nvidia-smi -i 0 --power-limit=160 && sudo nvidia-smi -i 1 --power-limit=175

 

[superducky]

Presumably the driver will let 'er rip until it needs to throttle up against the thermal limit if you don't set these?

Looking through the help for nvidia-smi for the first time in years, they've really done a much better job of documenting this than they used to!

nvidia-smi topo -m in particular is cool for SMP and Ryzen/Threadripper machines: (this is a single 1900x)

will@daystrominstitute:~$ nvidia-smi topo -m
        GPU0    GPU1    GPU2    GPU3    CPU Affinity    NUMA Affinity
GPU0     X      PHB     SYS     SYS     0-3,8-11                N/A
GPU1    PHB      X      SYS     SYS     0-3,8-11                N/A
GPU2    SYS     SYS      X      PHB     4-7,12-15       1
GPU3    SYS     SYS     PHB      X      4-7,12-15       1

 

[HayesK]

Presumably the driver will let 'er rip until it needs to throttle up against the thermal limit if you don't set these?

For sure on the older gpu, but beginning to suspect that the newest cards may be starting to throttle as they approach the power limit too.

My 3070Ti were pulling 245-265w with no power limit set and about the same with PL set at 300w. I reduced the PL to 250 on 2 cards and now appear to be running about 245w. Collecting HFM data to see how the ppd compares at various PL settings on same projects. The small atom count projects usually draw less power and boost higher than the big atom count projects.

In the past, have set the PL at the point where the boost clock begins to be affected. Thinking about reducing the older gpu even further to improve ppd/power ratio.

 

[dfonda]

Thanks for the help H.

 

[HayesK]

nvidia-settings can be used to set gpu/memory overclock/underclock and manual fan speed. Some factory OC GPU have required UC to be folding stable, even when new, but also later as they aged.

For folding, my gpu typically have gpu OC, memory UC and manual fan speed.

 

[superducky]

Hayes with the great info as always thank you!

 

[superducky]

For reference, here's where my 1080s and 1070 are sitting relative to their upper bounds so I think I'm okay without trying to nudge them down.