Follow up post, decided to build the 2x RTX PRO 6000 tower.
Posted by Signal_Ad657@reddit | LocalLLaMA | View on Reddit | 115 comments
Following discussions on my last post, decided to put the effort in and merge my two separate RTX 6000 towers into one. Parts list enclosed, would love tips if you have any, happy to answer questions if you have any. Here we go…
Build List:
Platform
• CPU: AMD Threadripper PRO 7965WX
• Motherboard: ASUS Pro WS WRX90E-SAGE SE (WRX90, EEB, 128 PCIe 5.0 lanes, dual 10GbE, IPMI)
• RAM: 128GB DDR5-4800 ECC RDIMM — 4× Samsung M321R4GA3BB6-CQK
Compute
• 2× NVIDIA RTX PRO 6000 Blackwell (96GB GDDR7 ECC each)
• 192GB total VRAM, x16/x16 PCIe 5.0, 500W cap per card
Case
• Corsair 9000D RGB Airflow (SSI-EEB, no fans included)
Power
• MSI MEG Ai1600T PCIE5 — 1600W 80+ Titanium, dual native 12V-2x6
• Dedicated 20A 120V circuit
Cooling
• CPU: Noctua NH-U14S TR5-SP6, dual NF-A15 140mm
• Front intake: 3× iCUE LINK RX140 MAX
• Top exhaust: 3× iCUE LINK RX140 MAX
• Rear exhaust: 2× iCUE LINK RX120 RGB
Storage
• Samsung 9100 PRO 8TB w/heatsink — PCIe 5.0 x4, 14,800 MB/s (OS, models, stack)
• 2TB SSD (scratch — Qdrant, datasets, embeddings)
Networking
• Dual 10GbE onboard (Intel X710)
CheatCodesOfLife@reddit
I don't suppose you'd know if this board works with 4 * 64GB + 4 * 48GB RDIMM sticks?
I've got the TRX50 (only 4 RAM slots) and upgraded from 192GB -> 256GB last year.
So I want to upgrade to the WRX90 and put all 8 sticks in there but can't find anything online about whether or not this would work...
kidflashonnikes@reddit
despite the asus QVL for this board - the most widely tested and positive reported RAM is the DDR5 ECC Kingston Fury 5600 RAM. ANything else really is a gamble
kidflashonnikes@reddit
I see lot of people saying no to the Noctua air cooler. I have 4 RTX PRO 6000s, but they are the maxwell, so they are blower style cards and I use the noctua air cooler and its just fine for me. I have the 96 core Threadripper CPU, with the noctua air cooler and it works fine. On top of that, I have 1 TB of the 5600 kingstom RAM (DDR5 ECC) all sitting on the wrx90 sage mobo (same as the OP).
I just want to point out that the air cooled systems are actually best for the CPU - in fact, many server grade builds use air coolers - and yes the flow of air is different in a server rack to maximize CPU performance comapared to a case, but none of you are doing this. On top of that, just get the Phanteks T30s - the reason these are special fans is because most fans are 25 mm - the phanteks T30 are 30mm wide, which means the extra 5 mm (cue size jokes) actually matter here - they push air into every single crevice because of this extra 5mm.
I have my 4 stacked RTX PRO 6000s (maxwells) all housed on one board in the phanteks server pro II TG case and it works perfeclt with the air cooler system I have for the CPU. I just want to explain one more thing. I actually work at one of the largest AI labs in the world - my set up is always on everyday - literally. For me, the only time it is off is when I do maintence (once a week), no more for 1-2 hours (cleaning, inspection ect). For me, air cooling was a no briainer - as AIO liquid systems are a major no no for systems running 24/7.
Technical_Split_6315@reddit
And after 40k you won’t be able to run a model better than the average Joe will be running for 20$/month pepehands
Own_Mix_3755@reddit
That avarage joe possibly does not care about privacy. Some workflows require data to never leave the office. Nobody will spend this amount of money for plqying games on it.
LA_rent_Aficionado@reddit
Correction, I still play games on my 348 GB of vram rig when I get bored
Own_Mix_3755@reddit
You absolutely can, but it wasnt probably your main driver to buy it.
LA_rent_Aficionado@reddit
Absolutely
AlwaysLateToThaParty@reddit
My rtx 6000 pro is the best gaming GPU. Ultra settings everything.
V0dros@reddit
TIL you can run games on an RTX 6000 pro. I assumed gaming drivers wouldn't work on workstation cards.
thrownawaymane@reddit
You can run anything. Many game designers use the video ram left over for debug.
I know a guy that has a 6000 for reasons and only uses it for gaming. Lol
AlwaysLateToThaParty@reddit
Is seen a person serve two gaming systems from a single CPU by simply connecting the monitors to the GPU, and having two VMs serve two different mouse/keyboard setups. Different rooms. Crazy. Length restrictions of course, but still...
I use qwen3.5 122b/10a heretic mxfp4_MOE btw. Takes about 75-80GB. For what it is needed for? Can't over-state its effectiveness. It has already paid for itself many times over. But I can still play BG3 at max settings while it is earning its keep.
thrownawaymane@reddit
Yeah I originally got into VMs to create a setup like the one that person made. It's definitely a bit of a challenge.
I only have the means for a 5090 and I know it will hold me back. And the niche I'm targeting is... sparsely populated on the business end but has plenty of buyers. But I can't afford a 6000 no matter how I do the math.
AlwaysLateToThaParty@reddit
It's a 96GB 5090 with more cores.
laterbreh@reddit
This is average joe thinking. Maybe he wants a tax write off, or wants self reliance and privacy for his business. Stop thinking in such small terms.
LA_rent_Aficionado@reddit
For real, a lot of hate goes around on here to undermine people with more disposable income. Granted, some people are just flexing so I get why they may feel animosity. I don’t sense OP falls into the latter category
LA_rent_Aficionado@reddit
If your goal is to just use it for local coding, absolutely. But when it comes to anything training or data generation I really see huge value in local in terms of saving time - you can quickly iterate on workflows and training, faster stand up of multi-model environments and testing. I have about 350gb vram across 8 GPUs and can do so many things at once it is great
IrisColt@reddit
What would be the retail price of a similar machine? Genuinely asking, I don't want to build it from scratch.
Signal_Ad657@reddit (OP)
It’s WAY cheaper to build. Safely over 30k, shipping would be a huge pain. This thing is big and might be over 150lbs when it’s done. Just simple math as I’m adding it up in my head it’s roughly 30k in parts and components. I’ve seen pre built systems like this sell for 40-50k.
IrisColt@reddit
I really appreciate the detailed explanation... thank you!
Hearcharted@reddit
https://i.redd.it/ftk0polqn0vg1.gif
EmptyNothing8770@reddit
Let‘s see Paul Allens RTX Pro 6000 Tower.
Hearcharted@reddit
🧐
NoFaithlessness951@reddit
Other people buy a car for that price
twinkbulk@reddit
This doesn’t depreciate as hard and banks actually count it as an asset for collateral in full value
iMakeSense@reddit
Could you ELI5 this?
twinkbulk@reddit
Your business builds up a credit profile, a business credit card, a line of credit, or equipment financing. You use that to buy the computer with an RTX Pro 6000 (20k). Now the LLC owns a tangible, depreciable asset on its balance sheet. You go to a lender and say “my business owns this equipment worth $X.” They appraise it, and because it’s a real physical asset the business owns free and clear (or with enough equity), they’ll lend you a percentage of its value — typically 50–80% for equipment. So on an $20k pc, you’re looking at $12K–$16K back in a new loan against it.
OverclockingUnicorn@reddit
Err hardware does depreciate, maybe not over the last \~6mo, but in general it does. Its not an investment and I doubt you'd get that large of a loan using a 2 GPUs as collateral.
twinkbulk@reddit
I didn’t say it didn’t, I said it doesn’t depreciate as hard, look at the price of a 3090 today vs on release 10 years ago. Banks have started realizing they’re extremely easy to offload and retain value, in recent years even going up in value, so yes they take them as collateral and with the system he has it’s easily 30k.
Blaze6181@reddit
Oh shit. Did not know this. 📝
syndorthebore@reddit
I sold my 2 RTX A6000 for more than I bought them for.
laterbreh@reddit
Other people buy these as a tax write off for their business.
timuela@reddit
Hmm... How would that work exactly?
Signal_Ad657@reddit (OP)
Just against revenue. If I make 25k, I’d usually owe like $10,000 in taxes on that income as a small business. If I buy 25k in hardware, I can put it against that income and now I have $0 of income. 25k came in, 25k went out, no taxes to pay on the 25k because the expenses offset the income.
Now there’s another side to this. You can’t just keep buying things to offset income because eventually you need to pay the bills and put food on the table and keep the lights on. But yeah. What would have been 15k income and 10k taxes becomes a 25k super computer and no taxes.
Ideally you purchase things as a business that can offset your future costs. So 25k comes in, 25k super computer gets made, that offsets future API costs, subscription fees, etc. Or you can rent out the compute. So you avoid the initial taxes and reduce long term overhead (maybe even add a bit of income).
That’s all probably a bit much for this thread but that’s the gist of it. You asked a good question so I wanted you to have an answer.
Clean_Hyena7172@reddit
My car cost the same as one of those GPUs in my country.
epicskyes@reddit
My car cost 2k I built my computer from scratch for around 24k now it’s worth 30k up 6k since rampocalypse.
Such_Advantage_6949@reddit
A car in my country costs 15x rtx 6000 pro. That make it justified for me then?
BobbyL2k@reddit
You better get 16x then. Power of two number of cards is ideal as well.
Such_Advantage_6949@reddit
Yea that is a good idea, let me sell my car… oh wait, i dont have a car..
AeroelasticCowboy@reddit
You can't buy a decent car in your country for less than $70,000 USD?
No_Night679@reddit
Some sells a car to buy 2x RTX Pro 6000 coz they have to learn. So I guess everyone has a different set of priorities.
RedHood31@reddit
What are you using this bad boy for?
Signal_Ad657@reddit (OP)
Fine tuning, AI video with LTX, hosting bigger stuff, doing research/testing/benchmarking, agent work.
RedHood31@reddit
Sounds really interesting, are you documenting what you’re doing? Like a github or a youtube channel, would like to actually see some specifics
Signal_Ad657@reddit (OP)
I haven’t been I’ll start doing it. I have a GitHub for my main project where I also share research here: https://github.com/Light-Heart-Labs/DreamServer
RedHood31@reddit
When you start please add all the links on your github would love to follow, thanks for sharing
Signal_Ad657@reddit (OP)
Will do
Aroochacha@reddit
I want to ask about the pricing on the memory but I don't want to run you through that trauma again.
It's a great work station. Don't be afraid of using an MOE model with the CPU + System ram handling part of the inference work.
Signal_Ad657@reddit (OP)
The memory pricing (especially now that I decided to go 256GB) was insane. Like 7k for server grade memory. I lucked out at least by having 256GB to sell from the old towers so I’m getting some offset / relief there. And yes! Can’t wait to try this.
SSOMGDSJD@reddit
Quality build porn. Do you plan to expand ram to fit glm5.1 and shit? Did you consider a sapphire rapids Intel build for meaningful CPU inference? Claude code pointed me in that direct whilst daydreaming about having big boy money the other day
Signal_Ad657@reddit (OP)
Yeah I shelled out after feedback and got the extra bank of RAM. I might be mistaken but I think that’s what you’d pick for CPU based inference which might be a different setup than 2x6000’s. I keep thinking about the CPU, and keep coming back to memory bus speeds being my actual bottleneck with this build if I spill into offload. The degenerate in me wants to go beefier I just haven’t found a good reason yet.
SSOMGDSJD@reddit
You're correct, saph rapids would be whole different direction than 2x 6000s. Sounds like you're optimizing in the right direction (as far as I know lmao) with the memory speed bottleneck. Need open opus for more hardware beef at this point
coolzamasu@reddit
why only 1600W power?
Your both blackwell itself can consume 1200W combined and then something for PC etc etc easily if you are at high load, it can be easily 1300-1400W.
shouldn't you have 2000W power supply just for safe side?
laterbreh@reddit
i run 3x rtx pro on a 20a 120v circuit at 300w each on a 1600w psu about to put a 4th in. Platinum/Titanium powersupplies dont give a fuck. Not worried about CPU consuming power when vllm is running.
In inference for single user i cant feel the difference of dropping the cards to 300w honestly. Eyeball math is about 10 to 15% loss. Acceptable terms for more expansion capability.
Norwood_Reaper_@reddit
Did you get the Max q or the regular workstation version of the cards?
laterbreh@reddit
two workstations and 1 max q
Signal_Ad657@reddit (OP)
The titanium PSU is tested and rated for 95% of full load, so unless I hit max CPU and max dual GPUs all at the same time it should be comfy. I was looking a 2000w options and just wasn’t happy with them and so got a 1600w that was actually built and rated to be able to stay loaded near full power.
Such_Advantage_6949@reddit
U can power limit it a bit, i am running mine at 450w, from my test it make no difference, especially for inference work
Signal_Ad657@reddit (OP)
Yeah I just realized, I have a spare 1200w platinum PSU from the old tower so I’ll definitely use that for the board / CPU / fans etc.
Such_Advantage_6949@reddit
Asus sage recommend to use identical psu though, it stated in the manual
Signal_Ad657@reddit (OP)
Smart catch, thanks!
Such_Advantage_6949@reddit
Well it might not burn thing out but if u alrd spent this much it is better be safe with the power. I use dual corsair 1600i for mine, on the same board but with 6 gpus
Signal_Ad657@reddit (OP)
Yeah I just bought a second 1600w PSU. Overkill now but if I add another card or two later it won’t be.
novelstalker@reddit
I have the same configuration but with 265K + 160G RAM + Seasonic T1600W. The biggest problem is the temp of the 12v 2x6 plugs. The temp of the plugs are pretty high. So I reduced it to 420W. Another reason I reduce the 2 6000PRO to 420W is that I use a maximum of 1530W\~ in agentic tasks.
Signal_Ad657@reddit (OP)
Awesome advice thank you!
LA_rent_Aficionado@reddit
FWIW I run my pros at 400w or at 500w at times, performance is still fantastic as the last 200w is diminishing returns - I imagine many others do too
finevelyn@reddit
Running that setup at above 1000W total is going to be pretty inefficient. I think 1600W is easily enough with a good margin to spare.
qPandx@reddit
what do you do for work bruh
brobits@reddit
he starts his day by not calling his coworkers bruh
qPandx@reddit
which braincells gave you the conclusion that we are coworkers?
darkkite@reddit
tbf they're not coworkers
Crafty_Top_9366@reddit
I don’t know it’s cool That you were able to do that but I think that the price just to talk to the computer is not good enough because even if you get 96 GB of virtual Ram, you really can’t load much larger models like Kimi Ke 2.5 for 2.7
Succubus-Empress@reddit
What is your address?
Different_Bumblebee3@reddit
So what do you plan to do with this?
Signal_Ad657@reddit (OP)
Fine tuning, AI video with LTX, hosting bigger stuff, doing research/testing/benchmarking, agent work.
LaCipe@reddit
tbh...once we get a model that can actually compete with opus, I will probably also drop some good money like that, unfortunately there is nothing yet, even despite all the problems that opus is currently facing.
Norwood_Reaper_@reddit
This was my plan too. Something local that competes with Opus that can fit on \~200GB of VRAM and I'm in. Maybe sooner...
Different_Bumblebee3@reddit
Gotcha 👍. Does all this compute power help with LTX even though LTX is a web app?
Signal_Ad657@reddit (OP)
Oh 100%. LTX is a web UI but the actual video generation model runs locally on your GPU.
Metalmaxm@reddit
Suggestions, take them as you see fit (again, none cynical European).
- Serious bottleneck 10GbE, if your pairing with 2 pc's as one for ai. Minimum 200Gb.
- Books on your right side - put the big tower there in between books and table. So dogs cant smash it.
- Mini but active cooler with fan in nvme.
// If there are no dust filters at your tower. Buy them via ebay, super cheap.
// No idea if your renting or this is your place (no need to answer). Outside bellow your windows and above, 4x or 8x solar panels 700\~ watt each. Of course if there is space.
Some day, you will also read and check out my post :D
Zwan_oj@reddit
1.6kw… PSU… Should of gone 2kw… you WILL be getting random shutdowns.
Such_Advantage_6949@reddit
This is very solid build
Orlandocollins@reddit
Nice! I also have 2 and I am jealous of your pro threadripper. I went regular threadripper and now wish I hadn't. Never thought I would consider 4 cards.
DataGOGO@reddit
HIGHLY recommended you return the AMD threadripper and MB and get the Intel XEON workstation CPU instead, MUCH faster at all AI workloads + AMX; Supports 8-12 channel memory. (8 64GB on 8 channels twice as fast as 4 128GB in 4 channels)
Also, Not enough power supply. Get the ASUS loki 1200w SFX power supply and use that to power your motherboard / CPU /RAM fans, etc, That case has a dedicated mount for the second power supply.
Your 20A circuit *should be* be enough.
HopePupal@reddit
dude's got two RTX PRO 6000s, the CPU in that thing is purely ceremonial. just needs to have enough PCIe lanes to not bottleneck PCIe P2P
DataGOGO@reddit
Incorrect.
HopePupal@reddit
care to expand on that?
laterbreh@reddit
I cant even begin to imagine that hes going to use llamacpp or do cpu offload when he can do tensor parallel with those two cards and blow the pants off anything llamacpp is capable of.
DataGOGO@reddit
Incorrect.
laterbreh@reddit
Care to elaborate? Cause I can tell you llamacpp is dogshit at any serious workload. Do you actually own these cards? Have you bench marked them in actual work loads? Or do you want to just give your oppinion then not back it up when someone pushes back.
DataGOGO@reddit
Agree with you on llama.cpp
Was not talking about CPU inference, or only cpu inference.
Yes, I do.
TRT / TRT LLM is FAR better on all Blackwell’s than vLLM/sglang; even on the SM120 consumer cards.
laterbreh@reddit
Thank you, now we have context. Appreciated.
Party-Special-5177@reddit
Wait, is AMD not the CPU king anymore? I had thought both AMD and nvidia collectively ate Intel’s lunch.
I know all of the HPC workstations I’ve seen until recently are all epyc builds but maybe I’m out of touch
DataGOGO@reddit
For AI workloads, (and any other memory performance sensitive workloads), Xeons have always been FAR faster than the AMD’s
LA_rent_Aficionado@reddit
Asus recommends you use the same model power supply for both under dual PSU in the manual, I did it with the Loki SFX for a while but eventually swapped after I read into it more
DataGOGO@reddit
that only matters if you are splitting 1 bus with 2 power supplies.
in this case, having 1 PSU for system, and 1 for GPU's, it makes absolutely no difference, you don't even have to use ATX power supplies for the GPU's, and just run a dedicated 12v supply like servers do.
TheyCallMeDozer@reddit
I'm just going to go pretend my 2x R9700's is as powerful with a 1b model...
MachinaVerum@reddit
I’ve done a similar build and I really recommend that you don’t air cool the cpu. Your gpus are dumping 1200w of heat in front of that coolers intake. I tried that with a 4u-m in my build and it was a disaster in terms of cpu temps. I switched to a silverstone aio set as intake and it solved the problem.
Signal_Ad657@reddit (OP)
Was going to use this guy for CPU: Noctua NH-U14S TR5-SP6
MachinaVerum@reddit
Don’t. It’s gonna sit there sucking in all that really hot air. Go for an aio and set it to intake.
LA_rent_Aficionado@reddit
Agreed, I’d set up the radiator in the fan slots right above the mobo, it makes getting the power cables in harder but well worth it for a cleaner look and better airflow, the slot to the right of the mobo is fine too
Signal_Ad657@reddit (OP)
Will do thanks!
laterbreh@reddit
Had simmilar issue. Got a 3000rpm server chasis noctua as an exhaust fan, dropped temps on everything by a substantial margin. CPU temps dropped 10 degrees. Its loud as hell at full speed, but when youre blasting minimax in tensor parallel at 100tps with these two cards, that loud speed makes me feel like im in a spaceship during inference. So i find it an acceptable tradeoff.
MachinaVerum@reddit
If it works, it works!
FullstackSensei@reddit
low-cost #budgetbuild #affordable #economic #inexpensive #economical
laterbreh@reddit
cope.
RoomyRoots@reddit
This sub really makes me rethink my life choiced because I couldn't justify this investment with the usage I have done so far.
inthesearchof@reddit
Very nice. Maybe 8 sticks of ram instead of 4?
Pixer---@reddit
Does P2P work on the wrx90 board ? I have a wrx80 board that doesnt
beepbeepimmmajeep@reddit
Don’t talk to me or my son again vibes
sacrelege@reddit
where did you buy the RDIMMs from?
FinalCap2680@reddit
If you can, leave more room for RAM upgrades.
Party-Special-5177@reddit
Excellent choice of hardware! Love the build, love the view, love the reference-books-as-decoration on your bookshelf lol
Idk if there are more choices now, but back when I made a similar build last year, there were basically 2 choices of mobo; the sage-se was absolutely the correct pick lol.
Damn thing shocked me with how heavy it was. Had never seen a board so packed with electronics that it actually had proper heft to it.
Signal_Ad657@reddit (OP)
DUDE THE BOARD! I’ve never seen such a heavy board in my life it’s nuts. It feels like it weighs almost as much as the power supply.
I want to try all kinds of stuff with it. Larger model hosting, fine tuning, AI video gen on LTX. I originally had the two towers just because I got lucky on a LAN center closing and it was a good deal but it was never the best use case for the GPUs I’ve just been making it work.
Really excited to put them into a machine that actually makes sense for them.
Echo9Zulu-@reddit
Cries in dual b70 shared 15a 120v