Build Advice What parts to choose for machine learning (3D medical imaging) ?

[jmitura]

Hello I seek advice on what to choose best.
My budget is around 1500 dollars.

I already have:
PSU
Hard Drives ,
CPU - Intel Ultra 9 285K
and graphics cards - 2 x RTX 3090

Now I need motherboard, cooling, ram and case .

Some background: I use my PC mainly for prototyping machine learning models in medical imaging, with quite complex code, one of main incentive for current hardware upgrade is that code compilation for my models take very long (up to 3 hours) hence currently i will leave my existing gpu and power supply but need better cpu and faster ram . Hovewer in sth like 2 years i will upgrade the GPU's to some newer version (and I still want to have 2 gpu's). I use Linux ubuntu.

I do not care how the pc looks like, it also do not need to be very quiet (I keep it in storage room and got cables through the wall to place where i have desk and work).
I suppose my RAM speed would be very important for those large compilation issues as I observe high ram usage, hence I am interested in cudimm technology, and need at least 96 gb of ram .

I want to have motherboard that supports 2 GPUs but frankly i do not know how important it is to push for 2x pcie5 - in my case models are always quite complex and they spend usually more then a second up to 12 for iteration on a gpu memory before i get next batch loaded , i know that in some cases in multi gpu training when one shares parameters between gpu communication is more frequent but hard to tell for me how important it is really.

I was looking into Gigabyte AI top z890 - it has 2x pci 5.0 ; and asus rog hero pci 5.0 and pci 4.0 plus seems to have better support for ram performance .

Also in order to utilize latest cudimms (I had seen cudimms sold up to 9200 MT/s) one would need to unlock extreme memory profiling on intel cpu - I wonder how safe it is - and also weather i need seperate cooling for ram .

Frankly I am software guy and get lost fast in this hardware complexities.
Thanks for help !

 

[Misgar]

i do not know how important it is to push for 2x pcie5

I assume you're trying to make your new system compatible PCIe5 GPUs when they become available and affordabe, but for the time being I'd stick to PCIe4 and not worry too much about what might be available in 2027.
https://www.pcguide.com/gpu/what-gpu-has-pcie-5/

need at least 96 gb of ram .

(I had seen cudimms sold up to 9200 MT/s)

Good luck getting 96GB RAM to run at 9200MT/s. RAM speed isn't everything. System RAM is important, but do you have enough video RAM on your two RTX 3090s? Some RTX A6000 cards come with 48GB VRAM.

need to unlock extreme memory profiling on intel cpu - I wonder how safe it is - and also weather i need seperate cooling for ram .

The faster you overlock RAM, the more likely it is to crash. Whether or not you need separate RAM cooling is a different matter. I have only one (ancient) system with a dual fan RAM cooler fixed over 4 hot DIMMs. All my other systems run fine without dedicated RAM cooling, but I use air coolers on my CPUs, so there is some air flow around the DIMMs and VRMs.

 

[jmitura]

Thanks for response @Misgar
1) yes you are correct I am thinking about 2027 weather i will need pcie5 then , as I would prefere not to change motherboard then, thanks for sharing your thougths about it
2) About CUDIMM I know 9200 is unrealistic - I meant that there are some fast CUDIMM out there considering how fast I should try to go having intel 285k
3) If it is about crashing I assume I would just need to dial it back and should be ok but what is the risk of permanently demaging my hardware
4) more vram and cuda cores would always be nice, but for now what is stalling my work is mainly compilation some cpu preprocessing, then gpu speed and then gpu memory - although I expect it to evolve with time , Hence I am sure I will need to get better gpu and my assumption for now is I will do it in 2 years (maybe a bit earlier if i will spend less now on upgrade)

Hovewer knowing that do you have some hardware recomendations?

Thanks !