Puget Systems print logo


Read this article at https://www.pugetsystems.com/guides/1974
Dr Donald Kinghorn (Scientific Computing Advisor )

Quad RTX3090 GPU Wattage Limited "MaxQ" TensorFlow Performance

Written on November 13, 2020 by Dr Donald Kinghorn

TL;DR: Can you run 4 RTX3090's in a system under heavy compute load? Yes, by using nvidia-smi I was able to reduce the power limit on 4 GPUs from 350W to 280W and achieve over 95% of maximum performance. The total power load "at the wall" was reasonable for a single power supply and a modest US residential 110V, 15A power line. At least look at the plots below.


The NVIDIA RTX3090 is a very capable GPU for accelerated compute intensive tasks like ML/AI model training. It offers a significant performance improvement over the last generation RTX20 GPUs. It also uses at LOT more power! The TDP Wattage for the RTX3090 is 350W. This is OK for high performance system builds with 1, 2, or even 3 GPUs. What if you want to push that to 4 GPUs?

Can you run heavy compute loads on Quad RTX3090 GPUs without using 2 power supplies and within the constraints of a US 110V 15A power line?

Yes! It's possible to sensibly make use of 4 RTX3090's. How? By using the NVIDIA System Management Interface (nvidia-smi) to apply power limits to the GPUs. I was able to achieve 95% of maximum compute performance with a GPU power limit of around 280W. That is a power draw similar to using 3 RTX3090's at full (350W) but with much better performance!


This testing was inspired by what NVIDIA and their partners do for some of their laptop GPUs. They have a "feature" called "MaxQ". That's an abuse of an aeronautical engineering term and not really appropriate but, hey, it sounds good and the basic idea is great! I set out to find "MaxQ" for the RTX3090 with the goal of finding a stable 4 GPU configuration with good performance per Watt.

The core idea of MaxQ is simple. The GPU performance increases rapidly as the current (Wattage) is increased from some minimal value. However, as the current approaches the max TDP, performance gains diminish significantly. You are trying to find the "sweet spot". You can define "sweet spot" in several ways; it could be where the derivative (slope) of the power vs performance curve has dropped below 1 (linear). You don't actually have to calculate anything because it is obvious from just looking at the power vs performances curves. I set my criteria for "sweet spot" to be 95% of the performance at full TDP wattage. It will be very obvious in the plots.

I had use of a testbed system with 4 RTX3090's with 2 power supplies and sufficient power "at the wall" to run them at full power under heavy load. I collected a lot of data and have some interesting plots. After a description of the test setup and procedure we'll dive into the plots.

Test system


  • Intel Xeon 3265W: 24-cores (4.4/3.4 GHz)
  • Motherboard: Asus PRO WS C621-64L SAGE/10G (Intel C621-64L EATX)
  • Memory: 6x REG ECC DDR4-2933 32GB (192GB total)
  • PSU1: EVGA SuperNOVA P2 1600W (80 PLUS Platinum)
  • PSU2: Super Flower Leadex Platinum 850W (80 PLUS Platinum)
  • NVIDIA 4 RTX3090s Gigabyte GeForce RTX 3090 TURBO 24GB Blower


  • Ubuntu 20.04 Linux (Server)
  • Enroot 3.3.1
  • NVIDIA Driver Version: 455.38
  • nvidia-container-toolkit 1.3.0-1
  • nvidia-smi
  • NVIDIA NGC containers
    • nvcr.io/nvidia/tensorflow:20.10-tf1-py3

Test Jobs

Testing Script

#!/usr/bin/env bash
# Generate data for maxQ plot



# setup environment variables, etc.
sudo nvidia-smi -pm 1 -i 0,1,2,3

for w in {100..350..10}

sudo nvidia-smi -pl $w -i 0,1,2,3
#enroot start ngc-tf1-10 python nvidia-examples/cnn/resnet.py  --layers=50  --batch_size=96  --precision=fp32 -i 120 | tee ${OUT_FILE}
enroot start ngc-tf1-10 mpiexec --allow-run-as-root --bind-to socket -np 4 python nvidia-examples/cnn/resnet.py --layers=50 -b 96 --precision=fp32 -i 180 | tee ${OUT_FILE}

ips=$(grep -E '^s+[0-9]+s+' ${OUT_FILE} | tr -s ' ' | cut -d ' ' -f 4 | sort -n | tail -1 ) 

echo "$w, $ips" >> ${TEST_DIR}/maxQ.data
exit 0

Note: I am using NVIDIA enroot to run the docker container image 20.10-tf1-py3 from NGC. I used enroot to import the docker image and create a local container "sandbox" named ngc-tf1-10. This script could also be setup using docker. Also, since I'm using enroot running in user space I really didn't need --allow-run-as-root, but you would need that with docker. --bind-to socket is also not needed in my case since I was using a single socket system.

Setting GPU power limits with nvidia-smi

To set the power limits nvidia-smi can be used. Here are the needed commands,

Set Persistence Mode

"When persistence mode is enabled the NVIDIA driver remains loaded even when no active clients, such as X11 or nvidia-smi, exist."

I did this to keep any changes to the GPU runtime settings from reverting to defaults after job runs had completed. This may be unnecessary(?) Must be done as root and it does not survive a system reboot (only available on Linux).

sudo  nvidia-smi -pm 1 
sudo nvidia-smi --persistence-mode=1

This should set all of the GPUs but I use the "-i" flag to explicitly specify the GPUs by id. For example `-i 0,1,2,3` for all 4 GPU's in the system.

Set GPU Power Limits

Setting the GPU power limit to 280W can be done with,

sudo nvidia-smi -pl 280 
sudo nvidia-smi  --power-limit=280

Please see the NVIDIA nvidia-smi documentation for details. It's a very powerful tool!

You can monitor power usage during a job run with, ( -q query, -d display type, -i GPU ID, -l 1 repeat every 1 second )

nvidia-smi -q -d POWER -i 0 -l 1 | grep "Power Draw"

Power vs Performance Plots

A number of jobs were run using TensorFlow 1.5 with a ResNet50 CNN training on synthetic data. The jobs ran for several minutes and data was collected for 1 to 4 GPUs using fp32 and fp16 precision. The plots convey a lot of information!

I used a simple model function to fit a curve to the data points to aid in visualization (it does not have the correct asymptotic behavior) Scipy curve_fit was used to find fitting parameters. For example,

def fcn(x,a,b,c):
    return a  + b*x + c*np.log(1+x)
popt32x1, pcov = curve_fit(fcn, watts, fp32x1)

array([-8.38597282e+03, -5.26458024e+00,  1.83816919e+03])

3090 MaxQ fp32

  • The plot shows power vs performance curves for 1-4 GPU's
  • The yellow lines are indicators for the 95% performance point for each curve
  • You can clearly see that performance gains decrease closer to TPD wattage
  • You can also see that the multi-GPU performance scaling going from 1 to 4 GPU's is excellent
  • It is also clear how much performance is improved by using 4 GPUs at restricted power vs 3 GPUs at full power.
  • There was some erratic performance at low power levels. In some cases wattage under 180 seemed to be ignored and the GPUs would run without the power limit. I had varied the power per GPU from 100W to 350W in steps of 10 but dropped the results from under 180W because of this.

3090 MaxQ fp16


  • These curves using fp16 (Tensorcores) are similar to the fp32 job runs

The following Amdhal's Law plot shows how well ResNet50 with TensorFlow is scaling for multi-GPU. This scaling holds up with the power limits.

3090 MaxQ Amdhal

This last plot shows the performance with 4 RTX3090's. The scale of this plot makes it easier to see the performance change vs wattage. It would be quite reasonable to run these GPUs at 250W instead of 350W and "settle" for 93% of max performance!

This plot also shows the total system power draw "at the wall" as measured with 2 "Kill A Watt" meters. (and 2 power supplies) This should give an idea what to expect for the total system current load.

3090 MaxQ Total Power

Conclusions, Caveats, and Unanswered Questions

I hope this testing gives you an sense of the feasibility of running a quad RTX3090 workstation. You could probably do this at full power by using 2 power supplies and an isolated 110V 20A power-line. (or 2 15A lines) Even better would be running on a 240V line. However, if you are considering this for use at home or at an office or lab that has restrictive power resources then you likely need to limit power for the GPUs or just do a build with 3 instead of 4 GPUs. You may not be able to have a full power circuit dedicated for one machine! Even if you can get the needed power it is worth considering lowering the GPU power limit to be more energy efficient!


  • This power limit solution is system configuration based (not firmware based). You have to ensure that the limits are set properly before you start running jobs.
  • A high quality power supply like the 1600W EVGA that was used in this testing is essential. That PSU has good circuit protection and will shut the system down on over-current situations. This is what I observed when trying to run 4 GPUs at full power and job load with just that single PSU.
  • You could trip circuit breakers on the power lines you are using. If you start up a large job and the power limit configuration was not set for some reason you might have problems if you overload the power line!

Unanswered Questions

  • Is this going to be reasonable in a "production" ready system build? I was running on an open-air testbed. I did not monitor temperatures. These Gigabyte blower style cards are excellent, but with 4 of them in a case they will be packed tightly and will require exceptionally good air flow through the chassis. Also well controlled ambient temperature would be needed. Remember you are considering power draw around that of a hair-dryer. There will be a lot of heat generated when the system is running a heavy job load.
  • Will limiting the power to the GPUs affect other tasks, like running the display or other things that put a load on the GPU? Probably. I ran my testing with an Ubuntu 20.04 server install without a GUI. I was connected remotely and didn't do anything else with the GPUs other than monitor them during the testing.
  • Can you do this on a MS Windows 10 system? I may try to do this on Windows. The NVIDIA driver for Windows does install nvidia-smi but it's not on your execution path, you will have to search for it. Also, Windows does not support "persistance mode" which may be needed(?).
  • The big question is -- Will Puget Systems offer this as a system configuration option? I will say no ... unless we can do the power limits in the GPU firmware. We are working are trying to arrange this but even if we are successful there could be issues with warranties, RMA's etc..

I can't say that I 100% recommend running 4 RTX3090's with power limits in a production system. But it is really tempting! If you want to do this you may have to order a system with 3 GPUs that is physically capable of supporting 4 and add the 4th GPU and do the configuration yourself.

I will probably do a follow-up post to this with instructions on how to setup a Systemd Unit file do do the power limit setting automatically at boot. I may also try this on Windows. One thing is certain, the NVIDIA RTX3090 is an amazing GPU! I do highly recommend this and also the RTX3080 for ML/AI work with a desktop system. The issues I had with drivers, CUDA libraries and frameworks in early testing are going away, as expected. I think one more update on software will take care of any lingering problems.

I you have questions or suggestions for a inclusions in a followup post let me know in the comments.

Happy computing! --dbk @dbkinghorn

Looking for a GPU Accelerated Workstation?

Puget Systems offers a range of poweful and reliable systems that are tailor-made for your unique workflow.

Configure a System!

Labs Consultation Service

Our Labs team is available to provide in-depth hardware recommendations based on your workflow.

Find Out More!

Why Choose Puget Systems?

Built specifically for you

Rather than getting a generic workstation, our systems are designed around your unique workflow and are optimized for the work you do every day.

Fast Build Times

By keeping inventory of our most popular parts, and maintaining a short supply line to parts we need, we are able to offer an industry leading ship time.

We're Here, Give Us a Call!

We make sure our representatives are as accessible as possible, by phone and email. At Puget Systems, you can actually talk to a real person!

Lifetime Support/Labor Warranty

Even when your parts warranty expires, we continue to answer your questions and even fix your computer with no labor costs.

Click here for even more reasons!

Puget Systems Hardware Partners

Tags: NVIDIA, TensorFlow, RTX30 series, Machine Learning

Does a 4x 3090 setup works with risers (all the GPUs on four separate riser cables)? I am asking this question because there aren't many 2-slot 3090 available out there. Except for the blower style GPU that you used, the only other option that we have is to go for water blocks, which many folks are not willing to do because of leakage concerns. The blower style Gigabyte GPU is a far rarer breed of the 3090 that is very difficult to spot in the wild, in an already endangered market of RTX 3000 cards. So, there is a burning question in the DL community who wants to build a 4x 3090 or a 7x 3090 setup, will all of them work if all the GPUs are attached to risers? I am personally very comfortable with modding cases, as I had done in the past. So I can mount the cards on separate locations on my case or can get a custom case made from the shop that can hold all the 3090 GPUs in a manner that do not restrict the airflow. I can also go for a mining rig. But the GPUs need to work with all of them on risers. It would be great if you can check this on your 4x GPU machine. Maybe you can write a follow up article on the thermal performance of the all riser setup vs. an all onboard setup. I have come across a lot of people on the internet who wants to know the answer to this question.

Posted on 2020-11-17 02:51:53
Donald Kinghorn

Yes, the RTX30's are still generally hard to get and Gigabyte is the only one making a blower style that I know of. I do highly recommend that blower card ...

You could do riser cables as long as you understand the potential issues (and it looks like you do)
You need really good quality cables as short as you can get away with. I believe you need full X16 electrical and I don't know about issues with cables and PCIe4 You can run the cards at v3 but ... some open questions there ...
The side fan cards are massive! Be sure to check and mockup a layout before you go shopping :-) They are also heavy! You would certainly need some clever chassis mods ... probably best to do a custom chassis if you can.
You will need to be able to deal with well over a KW of extra heat in the room housing the system. ... and those side mount card fans are not very directional ...

It would be a challenge to do the system. Of course I can't officially recommend you try it. But, if I was personally in a research group right now I would probably be talking with the machine shop folks about chassis supports for those cards :-) Performance is pretty impressive!

I wont be able to do the riser testing myself, and really I may be lucky to get hands on my test rig again for awhile. We are getting as many cards in as we can but there is lots of demand and supply is holding up system deliveries a bit.

If you do this, please send me a photo! :-)

Posted on 2020-11-17 17:40:55

Thank you for your reply. I will definitely share a photo of my build once it is complete, but in order to reach there you gotta help me out

Posted on 2020-11-18 11:17:17
Donald Kinghorn

OK looks like you can get PCIe4 x16/x8/x16/x8 on that board -- should be OK
It recommended to use at least 2x CPU memory as total GPU mem (because of buffering pinning ant such) so 128GB may be OK but as amazing as it seems 256GB would be recommended for 4 3090's

I know nothing about those risers. It would be best to look for v4 spec'd risers but they could be hard to find. You may be OK with what you are considering??? especially if your I/O to GPU demands are not too high. ... but I would consider an experiment that could fail
PCIe bus engineering can be finicky. The MD is probably OK but who knows about the GPUs you end up with?? ... I would be optimistic but mentally prepared for a fail.

PSU's are probably OK you can jumper one of them on for 2 of the GPU's

Physical mounting and cooling are probably going to be OK (within the constraints of having working PCIe risers) ... but you will be shocked when you first pop a 3 slot 3090 out of the box!

I don't have resources for testing either! I had use of that testbed config for a few days (most of the testing was done over the weekend) it had to be in a location with adequate power available and it was ultimately just in the way ( nearly all of our testing space and resources go to Windows application testing )

Posted on 2020-11-18 17:28:36

Thanks for all the heads ups. You are right about the fact that there are multiple points of failure. I don't have much idea about how much inter-GPU bandwidth my experiments consume (I mostly train ResNet like architectures). It is likely that PCIe 3.0 riser cables will bottleneck the system. But then again Tim Dettmers in his blogs, emphasized time and again that PCIe saturation is not common at all. Pinned memory helps ease out the PCIe transfers quite a bit. Linkup does make PCIe Gen 4 riser cables. I am yet to find a review of it though.

I can totally understand why you won't be able to conduct more tests with the 4x RTX 3090 setup that you used for this post. Speaking of this post, it is certainly one of the most comprehensive multi-GPU testing that I have come across the web. This post will go a long way helping researchers and production engineers in designing their systems. Thanks for all your efforts.

Posted on 2020-11-18 19:52:25
Donald Kinghorn

Thanks for your kind words! :-)

You should be OK with models like ResNet (I'm not sure about some of the Transformer networks,
will be testing that soon) There is no Peer-to-Peer over PCIe on the RTX30's so that all goes to memcopy from CPU but that's usually not much of a problem. (as Tim points out :-)

I think the weakest link in your build will be the risers. Get an extra one just in case. I really do think it will work! with good quality risers and very secure mounting for the GPU's ( they are big and heavy so will physically stress connection points )

Posted on 2020-11-18 20:21:52

Nvidia anounced a new desktop with 4 A100 gpus. They used a refrigerant cooling system and they claim it's maintenance free. Is water cooling completely out of the question for puget? EK water blocks seems to have a professional product line for workstations.

Posted on 2020-11-19 12:39:47
Donald Kinghorn

That sucker is beautiful! :-) I plan to do program qualification for Puget to offer them! We've got A100's coming and will be looking at the A6000 and A40 soon too.

In my reply to Dimitri above I mentioned doing dual PSU GPU workstations in the early 2000's Many of those were liquid cooled too :-) Puget Systems used to do a very nice liquid cooled workstation but we dropped it because of support issues and overall customer experience. It's not completely off of our consideration but if we did it again it would have to be more like a fully integrated closed loop type of unit.

The blocks and fitting from EK look very nice!

Doing a liquid cooled system is a really fun project but can be very challenging. We have people that know how to do it right but ... I think if I suggested it to our production folks some of them would say COOL! followed quickly by NO WAY :-) ... some of them would throw things at me :-)

Posted on 2020-11-19 18:05:21

Isn't it problematic to use 2 PSUs? The GPU draws power from the mainboard but also from its own power cable. If those are connected to different PSUs they might have slightly different voltages. Could this be a problem? After a quick google search,it might also be problematic if the power is out of phase. Any thoughts on that?

Posted on 2020-11-19 14:35:04
Donald Kinghorn

It's has been fairly common to do this. This is regulated DC 12V for both. But, I'm not an electrical engineer! ... however, in the early days of GPU computing I built dozens of systems this way and never had any issues. I used to use a custom replay to trigger the second PSU but it seems to be OK to use a jumper on the 2nd PSU too. A lot of the power supply companies make these that plug into the 24pin connector for this purpose.

However, I don't really recommend using 2 power supplies! That is part of the reason I worked on the GPU power limiting. That way you can use a single 1600W PSU. We use the excellent EVGA SuperNOVA 1600W and I'm sure there are other high quality units available. I would not skimp on the quality of the PSU. That EVGA 1600W would trigger a system shutdown with the current draw from the sys + 3 3090s under full load. I consider that a good feature :-)

The bottom line from the testing I did for the post is that you can use a single PSU with a reasonable full system power draw and still get ~95% of the performance of the GPU's I consider that a great result, it's significantly better than I had expected!

Posted on 2020-11-19 17:43:33
Donald Kinghorn

quick note for anyone seeing this today Nov 19th or tomorrow early ... I'll be on our live stream Friday at 1:00pm PDT
Labs Open Office Hour: Dr. Don Kinghorn Discusses The Latest in High Performance Computing https://youtu.be/9E2yW5qsB1E

I expect to talk about the quad 3090 ... it's totally informal so questions and discussion are very welcome!

I'll remove this comment after tomorrow

Posted on 2020-11-19 18:45:07