Max-Q or Not -- Input requested

Hardware Hub Build a PC
1

Hi everyone,

I am in the process of putting together a new build. It has a 9970X Threadripper (non-pro), 256GB RAM, 1600W PSU, and a Fractal Meshify 3 XL case. The case has 2x140mm fans on top (right above the CPU area), 1x140mm fan in the rear, and 3 x140mm fans up front for intake.

The video card is where I am undecided – am I better off with the Max-Q or the regular Workstaion version of the RTX Pro 6000. It will be going into Slot 2 of the motherboard (TRX50 AI-Top), and will be the only card in the case. The most LLM work that I see for it, would be maybe training work over the course of a few hours (or less) at a time.

I’m leaning towards the Max-Q version, but am concerned about how much noisier I hear that it is. I’m also thinking that if I were to ever want to sell, the Max-Q might be an easier sell , as most professional users would probably lean towards that (so they can fit multiple cards easliy into a case).

Anyway, what do you guys think? Is my thinking/consideration flawed? Any input would be greatly appreciated!

2

Well here is what I know.

Both cards can be set to 300w, at 300w the max-q has a slight flops advantage over the workstation, depending on the card, you will need to set the workstation card from 305w to 400w to match the max-q (most are probably in the 320-350w range). At 600w vs 300w the workstation has a pretty wide advantage.

On pure flops metrics the workstation card as a single gpu in a case is probably the winner.

Not sure anyone really knows what resell is going to be like, I imagine both the max-q and workstation cards will be fairly easy to sell when the time comes to it, the server card will probably have a much harder time finding a buyer due to the special cooling requirements.

The two advantages I see the max-q having over the other two variants is its cooling setup and lower power consumption makes it easier to fit 4 into a single workstation.

3

Thanks for the response. I’m definitely leaning towards the workstation variant as I will only be using 1 GPU, and be placing it in Slot-2 of the motherboard. That leaves a nice gap between the GPU and the CPU/RAM area.

4

I really love the Workstation card, and it’s very quiet most of the time. The thing is… I kind of want another now. :rofl:

I’m almost convinced that I could run two workstation versions with not that much issue, but I think the reality is they would be limited by thermals on long runs.

Do you need the extra speed of the workstation card? It’s diminishing returns when adding more wattage at some point, and adding a second card with more vram might be more important to you. If it’s quiet you are after, I think the workstation is still the choice.

5

For the stuff I’ll be doing, the 96GB of a single card will be MORE than enough; plus, I really do value the quieter option :slight_smile:

1 Like
6

If you are planning to install a single GPU, I think the 600W card is the obvious choice. From my experience, it has a much more effective cooling design than the Max-Q and will run cooler and quieter when installed by itself into a machine. The limitation of the 600W card is that it expects there are no airflow restrictions on either side of the card, so stacking many GPUs in a single chassis will lead to thermal throttling.

I have a 600W card installed in a a Meshify 3 case (not the XL variant) and the coil whine is more noticeable than the fan noise from the card. I am using the 3 stock 140mm fans that came with the case, and the CPU is using a 280mm AIO for cooling (just an Intel 265k) and even at 100% GPU utilization I see around 85C temps with GPU fan at 50% speed. The coil whine is more noticeable than the fan noise, and the case fans are louder than the GPU fan in my experience. Overall, I would consider this to be a quiet configuration - but not silent.

I have another machine with quad Max-Qs installed. I did just experiment with ramping up one of the fans on the Max-Q card and it stays pretty quiet until around 70% fan speed, and then rapidly increases in volume as the speed ramps up from there. Even at 100% a single Max-Q card probably won’t be louder than your case fans at 100%, but if noise is a concern you won’t be happy with it.

I did spin up just one of the Max-Q cards to 100% utilization for several minutes and it settled at around 85C and 65% fan speed. And that is for 300W TDP instead of 600W TDP.

One exception that would cause me to consider the Max-Q over the 600W card is if you are having difficulty keeping your CPU cool as-is. The 600W card exhausts into the direction of your CPU, and my CPU temps dropped around 15C when I swapped to Max-Q cards. (This was a 7965WX CPU that is air-cooled with an NH-D9). Although, honestly if this was a problem I would just upgrade the CPU cooler and still go with the 600W card for a single-GPU setup.

5 Likes
7

I’m using the Noctua NH-U14S TR5-SP6 cooler for the 9970X CPU, and so far cooling has not been an issue (but I’m still on my RTX 4090, which still exhausts out the back like a blower card would).

Here is what I was thinking with the Blackwell (Workstation edition): Since it blows GPU air upwards towards the CPU/RAM area, I figured I would put the card in Slot-2 — this would give a fairly good gap bewteen CPU an GPU. Then with the three 140mm front intake fans, single 140mm exhaust fan, and two 140mm at the case top (over the CPU/RAM area), I would be hoping that the amount of hot GPU air impacting the CPU/RAM would be mitigated enough to keep things from getting too hot.

8

I’ve seen in some tests that Max-Q GPUs can be up to 100% slower with large LMs. Have you noticed the same difference between your Max-Q and your workstation GPU?

9

I have been captioning images using Qwen3-VL, so I ran some of my data set through a single Max-Q GPU @ 300W and also a single 600W GPU.

The job has a system prompt with instructions on how to caption a photo, and the user prompt consists of the image. Thinking is enabled, and it responds with a paragraph or two describing the image.

This is not a benchmark, but here is the stats output from vllm for the first few minutes of handling the requests after launching.

My command for vllm was:

vllm serve Qwen/Qwen3-VL-32B-Thinking --reasoning-parser "deepseek_r1" --max-model-len 64k \
  --limit-mm-per-prompt.video 0 \
  --async-scheduling \
  --gpu-memory-utilization 0.93 \
  --max-num-seqs 128

The stats vllm posted with the Max-Q card @ 300W:

Engine 000: Avg prompt throughput: 2846.4 tokens/s, Avg generation throughput: 2.5 tokens/s, Running: 17 reqs, Waiting: 77 reqs, GPU KV cache usage: 37.0%, Prefix cache hit rate: 8.9%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 3557.9 tokens/s, Avg generation throughput: 8.6 tokens/s, Running: 32 reqs, Waiting: 96 reqs, GPU KV cache usage: 71.4%, Prefix cache hit rate: 9.1%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 3558.0 tokens/s, Avg generation throughput: 14.7 tokens/s, Running: 44 reqs, Waiting: 84 reqs, GPU KV cache usage: 99.7%, Prefix cache hit rate: 9.2%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 474.4 tokens/s, Avg generation throughput: 455.3 tokens/s, Running: 41 reqs, Waiting: 87 reqs, GPU KV cache usage: 97.7%, Prefix cache hit rate: 9.3%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 433.6 tokens/s, Running: 40 reqs, Waiting: 88 reqs, GPU KV cache usage: 99.7%, Prefix cache hit rate: 9.3%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 421.4 tokens/s, Running: 38 reqs, Waiting: 90 reqs, GPU KV cache usage: 99.0%, Prefix cache hit rate: 9.3%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 349.8 tokens/s, Running: 37 reqs, Waiting: 91 reqs, GPU KV cache usage: 99.9%, Prefix cache hit rate: 9.3%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 309.5 tokens/s, Running: 36 reqs, Waiting: 92 reqs, GPU KV cache usage: 99.4%, Prefix cache hit rate: 9.3%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 948.8 tokens/s, Avg generation throughput: 251.5 tokens/s, Running: 36 reqs, Waiting: 92 reqs, GPU KV cache usage: 99.3%, Prefix cache hit rate: 9.4%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 1186.0 tokens/s, Avg generation throughput: 247.9 tokens/s, Running: 36 reqs, Waiting: 92 reqs, GPU KV cache usage: 98.5%, Prefix cache hit rate: 9.4%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 1186.0 tokens/s, Avg generation throughput: 232.2 tokens/s, Running: 37 reqs, Waiting: 90 reqs, GPU KV cache usage: 99.7%, Prefix cache hit rate: 9.4%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 1186.0 tokens/s, Avg generation throughput: 257.5 tokens/s, Running: 37 reqs, Waiting: 91 reqs, GPU KV cache usage: 99.3%, Prefix cache hit rate: 9.4%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 711.5 tokens/s, Avg generation throughput: 280.8 tokens/s, Running: 37 reqs, Waiting: 90 reqs, GPU KV cache usage: 98.8%, Prefix cache hit rate: 9.4%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 948.7 tokens/s, Avg generation throughput: 306.8 tokens/s, Running: 37 reqs, Waiting: 91 reqs, GPU KV cache usage: 99.3%, Prefix cache hit rate: 9.4%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 474.3 tokens/s, Avg generation throughput: 326.2 tokens/s, Running: 36 reqs, Waiting: 92 reqs, GPU KV cache usage: 98.5%, Prefix cache hit rate: 9.4%, MM cache hit rate: 0.0%

Stats posted with 600W card:

Engine 000: Avg prompt throughput: 2846.4 tokens/s, Avg generation throughput: 2.5 tokens/s, Running: 17 reqs, Waiting: 111 reqs, GPU KV cache usage: 37.0%, Prefix cache hit rate: 8.9%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 4506.8 tokens/s, Avg generation throughput: 11.7 tokens/s, Running: 36 reqs, Waiting: 92 reqs, GPU KV cache usage: 80.0%, Prefix cache hit rate: 9.2%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 3083.2 tokens/s, Avg generation throughput: 303.2 tokens/s, Running: 42 reqs, Waiting: 86 reqs, GPU KV cache usage: 98.2%, Prefix cache hit rate: 9.3%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 609.4 tokens/s, Running: 40 reqs, Waiting: 88 reqs, GPU KV cache usage: 99.9%, Prefix cache hit rate: 9.3%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 574.3 tokens/s, Running: 37 reqs, Waiting: 91 reqs, GPU KV cache usage: 98.2%, Prefix cache hit rate: 9.3%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 396.9 tokens/s, Running: 36 reqs, Waiting: 92 reqs, GPU KV cache usage: 98.6%, Prefix cache hit rate: 9.3%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 711.6 tokens/s, Avg generation throughput: 363.7 tokens/s, Running: 36 reqs, Waiting: 91 reqs, GPU KV cache usage: 98.8%, Prefix cache hit rate: 9.3%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 2134.7 tokens/s, Avg generation throughput: 295.2 tokens/s, Running: 37 reqs, Waiting: 90 reqs, GPU KV cache usage: 98.8%, Prefix cache hit rate: 9.4%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 1423.2 tokens/s, Avg generation throughput: 376.4 tokens/s, Running: 37 reqs, Waiting: 91 reqs, GPU KV cache usage: 99.1%, Prefix cache hit rate: 9.4%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 1897.2 tokens/s, Avg generation throughput: 324.2 tokens/s, Running: 38 reqs, Waiting: 90 reqs, GPU KV cache usage: 98.7%, Prefix cache hit rate: 9.4%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 1185.9 tokens/s, Avg generation throughput: 395.2 tokens/s, Running: 38 reqs, Waiting: 90 reqs, GPU KV cache usage: 99.2%, Prefix cache hit rate: 9.4%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 711.6 tokens/s, Avg generation throughput: 450.5 tokens/s, Running: 37 reqs, Waiting: 91 reqs, GPU KV cache usage: 98.7%, Prefix cache hit rate: 9.4%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 948.8 tokens/s, Avg generation throughput: 406.3 tokens/s, Running: 37 reqs, Waiting: 91 reqs, GPU KV cache usage: 99.7%, Prefix cache hit rate: 9.4%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 711.5 tokens/s, Avg generation throughput: 428.8 tokens/s, Running: 36 reqs, Waiting: 92 reqs, GPU KV cache usage: 99.4%, Prefix cache hit rate: 9.4%, MM cache hit rate: 0.0%
Engine 000: Avg prompt throughput: 1660.3 tokens/s, Avg generation throughput: 300.6 tokens/s, Running: 36 reqs, Waiting: 92 reqs, GPU KV cache usage: 97.8%, Prefix cache hit rate: 9.4%, MM cache hit rate: 0.0%

Notably this is processing multiple requests simultaneously, so the increased compute speed of the 600W card will be put to use. If you are running things in LM Studio or Ollama I think you are only able to process a single request at a time, so there won’t be as much of a performance delta due to memory bandwidth being the bottleneck (and VRAM is the same speed on both cards).

Just glancing at the numbers, the Max-Q card seems to be about 75% of the speed for Qwen3-VL-32B (BF16)

10

Hello there NewBuild,

As a newby with some chops, I have been down the AI motherboard and video card rabbit hole. I think Ive seen every youtube video and configured on every builders website numerous times. This may be old news by now, but I just figured these things out - I told you I was a newby.

Puget Systems has figured this all out. Its in their standard builds.

The X870E build are for one GPU - the RTX 6000 Pro or Max-Q. But you should just get the Non-Max Q if you only want one card.

The TRX50 can handle 2 RTX Pro cards but THEY HAVE TO BE THE MAX Q. Makes sense as 600 watts x 3 = 1800 watts! That would cook your breaker panel.

The WRX 80 will handle 3 Max Q’s.

Dont hop on the Upgrade Analysis Paralysis treadmill unless you want to spend two months of learning.

The plus side of the Max Q over the smaller Video Ram cards is you dont have to buy as much motherboard Ram. 64 Gig will handle most of the quantized models, as most of the programs you will select the right quantization to run fully in the 96gig vram.

Also remember, how much time do you want to spend either quantizing the models or just trying to decipher the hugging face website.

If you are serious about running AI models locally, jump off the MaxQ cliff and get one.

Another plus is you can use the AI rig to analyze all your personal invoices and cancelled cheks for the past years and learn a lot about yourself.

I still havent boought a rig. US$9.000 is still a hell of a lot of money for a video card!

Please be easy on me Wendell, remember I am a newby!

11

So, I built this and it’s great with two 600w workstation cards. I spent a whole lot on it though.

I don’t regret any decisions, but living in the US, 1600w PSU seems dicey. I don’t have a 220v outlet in my whole rental house so I can’t even find an UPS that will work with this rig.

In short, I probably should have gone max q, and I’m now limited to what I have and can’t add any more cards, even if I wanted to at this power limit.

P.S. I didn’t even realize I already posted before hah, I guess you can see the evolution of my thoughts.

12

You can tune the TDP down if you need to. Maybe I missed part of the conversation. :joy:

1 Like
13

Yes, I can turn it down, but really I wanna get turnt up?

It’s fine as is right now, but I can’t expand further…

Looking for colos in the Bay Area actually, might be better to transfer to a rack in a joint that isn’t a (>100 year old) house.

1 Like