RTX 6000 PRO Blackwell Max Q? Non Max Q?
Posted by Opening_Progress6820@reddit | LocalLLaMA | View on Reddit | 30 comments
Hello everyone,
I’m looking for some advice on upgrading my personal GPU server for research purposes. I’m considering the RTX 6000 PRO Blackwell, but I’m currently debating between the Max-Q and non-Max-Q versions.
From what I understand, the Max-Q version operates at roughly half the power and delivers about 12% lower performance compared to the full-power version.
My question is this:
If I manually limit the power of the non-Max-Q version to the same level as the Max-Q, would the performance be similar, or could it be better than the Max-Q by more than 12%?
My reasoning is that the non-Max-Q version might be more efficient at lower power levels due to better thermal and power delivery design, even when underclocked.
Has anyone tested this or seen benchmarks comparing the two under the same power limits?
Thanks in advance!
MengerianMango@reddit
nvidia-smi says the min power on the non Max Q is 400. On Linux, I dont see a way to set it to 300.
Get the non Q if you're only going to run one gpu. Get the Q if you're going to run multiple.
Opening_Progress6820@reddit (OP)
Appreciate you sharing your experience. I've been looking everywhere for info like this!
MengerianMango@reddit
Yeah I'm wrong. I had a 5090 before this (returned it) and I misremembered. It can go as low as 150. My bad bro
MelodicRecognition7@reddit
are you sure it was 5090 and 150W? I've just read in another thread that 5090 is limited to 400W minimum.
it seems that jacket is messing with the drivers, perhaps installing older version would help.
MengerianMango@reddit
So the confusion in this comment chain of mine was that I had a 5090 and it had a minimum of 400w. I took it back to MC. Got a pro 6000. Then I misremembered/mistakenly assumed that the 6000 had the same min wattage.
5090 min: 400 6000 min: 150
Opening_Progress6820@reddit (OP)
No problem! Glad to know about this.
sob727@reddit
Parent is wrong. It can be set lower than 300W actually. Maybe parent has a software issue.
MelodicRecognition7@reddit
the info is wrong though
MelodicRecognition7@reddit
the non-Max-Q version definitely could be powered down below 400W: https://old.reddit.com/r/LocalLLaMA/comments/1kvf8d2/nvidia_rtx_pro_6000_workstation_96gb_benchmarks/
try to update drivers
MengerianMango@reddit
That guy is on windows. There are a few things you can't do on Linux. I think this is one of them. Also can't overclock or undervolt afaik
I wasn't sure if the 400 limit was specific to Linux or not, so that's why I predicated my statement the way I did. Good that you added extra context tho. You should ping OP
vibjelo@reddit
You can definitely both overclock and undervolt (setting the PL) on Linux with nvidia cards.
MengerianMango@reddit
Can you show me how? Sincere question
vibjelo@reddit
Works perfectly well for me and sets the power limit to 300 as expected. Lots of information about it here: https://wiki.archlinux.org/title/NVIDIA/Tips_and_tricks#Overclocking_and_cooling
DAlmighty@reddit
This isn’t true.
vibjelo@reddit
The choice basically comes down to how many cards you want to use right next to each other.
The "normal" version has normal open-air fans, that within the chassis both takes air and outputs air still within the chassis. This works fine when there is one, but if you have two next to each other, one would output their hot air straight into the other
The Max Q version however has a different fan setup (and lower power limit) where instead it pushes the air out of the chassis, so having them next to each other works fine. But only makes sense if you're actually planning to have more cards, or if you expect to not be able to upgrade the PSU has it's a lot easier to fit 600W (300+300) than 1200W (300+300).
TLDR: More than one card next to each other? Go for Max Q. Otherwise the "normal"/workstation edition.
Opening_Progress6820@reddit (OP)
Oh, I didn't realize that cooling part. Makes total sense now. Then for me, non-Max Q is the choice. Thank you so much for your detailed explanation!
No-Consequence-1779@reddit
Cooling matters more if your running long processes. For 5-10 minute inferences vary 30 minutes, it doesn’t matter too much. Really only when training. Then of course you can change and upgrade fan configuration.
MelodicRecognition7@reddit
what about taping an extra cooler at the side of both cards so the hot air would be pushed to the back of the chassis? i.e.
vibjelo@reddit
No worries, I saw some others briefly touched on it already but didn't properly explain so glad it helped!
Thalesian@reddit
I posted here yesterday with some thermal images that might be helpful. I don’t have a Max-Q but an older Ada Lovelace RTX 6000, but the cooling system is the same. My impression is that the cooling solution they designed for the 600w workstation edition is completely superior to the older blower style. We’ll have to wait for benchmarks, but I don’t see any value added for the Max-Q if you are talking about a single card in a workstation. Max-Q makes much more sense if you want to stack multiple cards on a PCI 5 mother board as their blower design is designed to work in tandem with other cards.
Long story short, if you are getting one gpu, get the 600w workstation version and don’t worry about power limitations if your PSU can support it. If you plan on having multiple cards, then the Max-Q is a superior choice. Make sure to measure things out - 600w is 12 inches wide, while Max-Q is 10.5.
Mr_Moonsilver@reddit
Wendell from level 1 techs made a video on YT about this exact question
sob727@reddit
Yes good video. OP this should help you.
Herr_Drosselmeyer@reddit
Performance should be identical with the same power target, though you may not be able to set the regular version that low.
I think it's more about the cooling solution's design than the max power. The regular version is a flow-through design that works well in a regular PC case, just like a 5090 does. If you're going to use just one card, two at a push, especially in a desktop computer, that's the one to get.
The Max-Q is a blower design, meant to be used in multi-gpu setups, since they exhaust towards the I/O side, rather then up/down. This makes stacking them in a case much more viable. The regular version cards would end up exhausting towards each other in such a configuration, an obvious isse for thermals. If you're going for a multi-gpu setup, get the Max-Q.
sob727@reddit
Yes you can set the regular version that low. Even lower.
Rich_Repeat_22@reddit
There are benchmarks between the two versions.
Imho given the same price both have, get the full power version and undervolt it to your needs.
serendipity98765@reddit
Did the reviews come out?
JohnnyOR@reddit
If it helps, if you go for the non-Max Q the cooling looks a lot like that of the 5090, so it works if you have only one in a workstation. We got a 2U form factor server this week that has 2, but we needed to spec the Max Q to make the thermals work
HomeWinter6905@reddit
Let us know your findings please, following.
MelodicRecognition7@reddit
I haven't seen the benchmarks but I've asked a somewhat similar question recently - why would I want to buy a low power version when I could just buy a "full power" version and power limit it to 300W, and got a very valid answer: the low power version cards are intended to be used in a multi-GPU setup, multiple cards stacked in one computer case, so if you plan to use just one card then go with the full power version. And at some point you might want to get these additional 12% for extra 300W power draw lol
bullerwins@reddit
Try to go into the level1 forums as people there have those cards and are doing experiments.