Recommendations on new 2 x RTX 3090 setup

Sometimes people like having the ability to upgrade parts of their system. With EOL sockets you’re not going to be able to really upgrade your CPU later. That being said, it should work just fine and if you’re getting a good deal on the parts then I’d say go for it.

The only immediately apparent difference I saw was one had 10G networking. I’m guessing they are probably pretty similar otherwise. 10G networking is only advantageous if you have other 10G devices that allow you to take advantage of the faster speeds.

Then i could upgrade cpu later for SAGE PRO 2 but not for SAGE ?

I might get SAGE motherboard from ebay for $500 but pro is around $850 from walmart. The one from ebay is opened but not used though (as per the seller). Is it safe to buy from ebay or not worth it as not a good deal (because i might not get 3 years warranty for ebay mobo) ?

Do you see any issue with this case for dual rtx 3090 ti :Lian Li O11 Dynamic XL ROG Tempered Glass eATX Full Tower Computer Case - Black - Micro Center

No, they are both X299.

It’s definitely riskier to buy from eBay, especially used, however I still do from time to time when I personally feel the risk/reward is good enough or when the thing I want to buy is no longer made. My guess is the listing you’re seeing on Walmart is a marketplace product which may also mean it’s used with no warranty. For the prices you are listing this doesn’t seem like it’s a lot cheaper than a new system.

Should be fine. Lambda labs uses this case for dual 3090/4090 machines. I have used with a dual 3090 machine but that was water cooled.

@matdmiller Thanks for the information. I am always confused but could not verify it as why not people use riser cable (and I have not seen a lot of example using that and @balnazzar mentions that it might reduce performance. Can you run an experiment with and without the riser cable or if you are aware of such experiments in deep learning (for other applications people tend to claim that does not affect much) ? …Because if riser cable does not degradfe performance much people could use rack mounted system to add 3 or 4 gpus ?

I don’t remember running across any benchmarks. I was at one point thinking of using a riser cable after a few slots stopped working on my motherboard so I bought one and then figured out there’s no easy place to put my GPU in my case while the GPU was attached to the riser so I ended up just replacing the motherboard.

I would try and find a case that is designed for this before going too far unless you’re comfortable designing and fabricating something to make this work yourself.

Also, I would not recommend having more than 2 GPU’s (for cost/benefit reasons) unless you already specifically know how you’re going to use more than that. 2 is nice because you can be doing a longer training run on one GPU while doing development and smaller experiments on the other. That being said, most people in the course learn purely using cloud compute with a single GPU. I am going on the assumption that you don’t specifically have a need for that many GPU’s and you’re just looking for recommendations for a starter machine to learn deep learning with. I started off with a machine capable of 4x GPU’s with 3 actually installed and found that I didn’t actually utilize the 3rd GPU very often. If I were to do it again, I would go with a dual GPU system.

Thanks @matdmiller . I finalized two PSU but both are looking alost similar one being of 2014 and other being of 2017. But it seems nothing has changed much in 2022 for both.

  1. https://www.newegg.com/evga-220-t2-1600-x1-1600w/p/N82E16817438041?Item=N82E16817438041

  2. https://www.newegg.com/corsair-ax1600i-cp-9020087-na-1600w/p/N82E16817139226?Item=N82E16817139226

1 is $480 though, but 2 ($610) uses some Gallium nitride:The Next Generation of Power | Navitas. Any idea ? … Or i could go for a lot cheaper one ($363)[EVGA SuperNOVA 1600 G+ 1600 Watt 80 Plus Gold ATX Fully Modular Power Supply - Micro Center] … as i might add one or two more gpu in future so just wanted to leave that possibility ?

I think both EVGA and Corsair are reputable brands and 1600W is pretty much the biggest consumer grade PSU you can get. I would buy whatever is cheapest overall. The more expensive ones are more efficient. You could calculate potential cost savings based on your anticipated usage, electricity costs and the efficiencies for different PSUs. You can look up tested efficiencies here: 80 PLUS® Certified Power Supplies and Manufacturers | CLEAResult .

One thing to note is there is a new PSU specification (ATX 3.0 / PCIE5) that came out in 2022 but unfortunately there aren’t many PSU options that meet the new spec and I don’t think there are any high capacity PSUs that meet it yet. The 4090 supports (but does not require - comes with an adapter) the new spec. At some point GPU’s probably will require the new spec, but it’s hard to know when that will actually be. Just something to consider - when you do a GPU upgrade in the future you MAY also need to do a PSU upgrade so that may factor into your ‘payback period’ calculation/how long you may actually keep/use this PSU.

I got super lucky and got this on sale for $200 Amazon.com which was an upgrade from my 1,000W PSU I had before upgrading to 3090s.

Well in that case, it makes more sense to go for $330 PSU (instead of $500 or $600) https://www.amazon.com/gp/product/B08F1DKWX5/ref=ppx_yo_dt_b_asin_title_o00_s00?ie=UTF8&th=1 ?

I decided to go for Fractal Design Define 7 XL case (Robot or human?) instead of (Lian Li O11 Dynamic XL ROG Tempered Glass eATX Full Tower Computer Case - Black - Micro Center). I hope that should not be a problem with that case ?

I don’t see anything wrong with it. It’s a full tower case so it should fit an EATX motherboard and multiple GPU’s. I would just do a little research on how well it does with cooling/airflow to make sure it can keep your air cooled GPU’s cool. Others may suggest if you need to add fans to the case or not.

Hey, I’m currently building out a machine for learning ML and playing around with models. I also need to be able to use it whilst training is ongoing, so I’ve locked down the following parts list:
2x 3090 FE (2 slot)
13900K
990 Pro NVME
2x32 Trident Z5 @6000mhz DDR5
HX1500i PSU

All on a custom loop with NVLINK between gpus.
What I’m struggling with is the motherboard and case. For the motherboard, I’m trying to work out whether running the gpus at 4.0x8 will have a significant performance impact. And if not, what boards support ddr5 ram, 2 gpus in the x16 slots, all whilst not bottlenecking either the gpus or the ram. Do you have any rec’s, I’m sick of trying to parse marketing BS about support.
For the case, I’m just looking for something that can handle the aforementioned hardware, a full loop, a couple of HDDs and the rad’s required to cool it all. The rads are the difficult part, I can’t find info about cases that at a minimum will suppport 2x360mm and 1x420mm (or ideally 3x420mm).

Any help you could provide with any of that would be really appreciated. Sorry for the wall of text everyone.

That is a tall order to want to support all of the radiators. The Fractal case mentioned above is one of the biggest out there. Assuming you can satisfy fan/rad clearances, it looks like you can run 480x120 up top, with 420x140 in the front, and 240x120 on the bottom, assuming the depth of the PSU does not interfere.

Yeah, that’s what stumps me. But people have been doing dual gpu builds for eons now, I must be missing a trick on how to cool everything. I calculated ~1100w under load, so 9x 120mm fans should be enough.
Thanks for the fractal rec, I’ll look into it. Might just go MoRa if it proves too frustrating to find a case.

according to EKWB, you can cool the cpu/ram/2xGPU with just 2x420rads

I suspect that it will not have a significant impact if any at all. You can read more about that here The Best GPUs for Deep Learning in 2023 — An In-depth Analysis

It’s annoying, but the only way to really know how the PCIE is set up is to read the manuals. The vast majority of people (gamers) only do 1 GPU which is what the marketing is tailored for.

I don’t have any experience with this motherboard, but it does support splitting the 5.0x16 bandwidth evenly between the top 2 x16 slots which is what you’re looking for. Many of the current gen Intel CPU motherboards do not do this - some only give you 4.0x4 that has gone through the chipset - you don’t want this. This case would allow you to do air cooled cards as there is 4 slot spacing between the 2 slots.

FE’s are not 2 slot, I assume you mean they are 2 slot after adding a custom water cooling block to them? I would suspect they would be 1 slot if that was the case. Or do you mean they are hybrid cards?

I’m guessing that you have more cooling here than is needed. I have EVGA 3090 hybrid cards with 240mm rads and they run pretty cool, especially compared to some of the numbers I’ve seen on air cooled equivalents.

I built a PC for work with this case. It has a TR pro with a 360 aio (I would have gone with 240 but couldn’t find any in stock) and 2x3090 hybrid cards w/ 240’s. I had to 3D print a front panel with rad opening to get all of the rads to fit in this case. It is one of the largest cases available. The 360 is on top, and there is a 240 on the back/(right side when facing front of case) and a 240 in the front panel which I 3D printed to replace the stock glass panel. There are 3x120mm fans on the bottom of the case and a rad would probably fit there, but since all of my rads are aio/hybrid they can’t be below the pump. If you look at the mfg website for this case they say you can fit 3x360 rads on a custom loop which makes it ok to put one on the bottom of the case. There is plenty of room for hard drives too.

2 Likes

@matdmiller ,
How do you control fan or pump speed in linux/ubuntu (as there is no support for msi or evga pump speed in linux but only supported in windows) ?

Have water cooling and my pump/fans are regulated in BIOS… x570 PRO mobo

1 Like

I’ve only used AIO/Hybrid cards with default OOTB settings. I’ve never had to set anything.