RTX 5090 + PhD work — what non-gaming hardware should I pair it with?
Posted by Alternative_Art2984@reddit | buildapc | View on Reddit | 17 comments
Hi everyone,
I’m a PhD student (in Australia) working on deep learning / video understanding, and I’ve recently got my hands on an RTX 5090 (24 GB). I’m not building this system for gaming at all purely for research workloads (training models, large datasets, possibly multi-GPU in the future).
I’m trying to figure out the rest of the setup, but most recommendations online are very gaming-focused, which isn’t really helpful for my use case.
Right now I’m unsure about:
- CPU choice (should I go high core count like Ryzen 9 / Threadripper, or something else?)
- Motherboard (reliability + PCIe lanes matter more than RGB or gaming features)
- RAM (capacity vs speed — is 64GB enough or should I go 128GB?)
- Storage (best setup for fast dataset loading — NVMe vs multiple drives?)
- PSU and cooling (for long training runs, stability is key)
My priorities are:
- Stability during long training jobs
- Fast data loading / preprocessing
- Scalability for future upgrades (maybe more GPUs later)
- No unnecessary “gaming” features
Budget is flexible but I don’t want to waste money on things that don’t actually improve research performance.
pmjm@reddit
PCIe lanes are determined by the CPU and not the motherboard. If you ever anticipate wanting to add more than one GPU and really take full advantage of it, you'll need Threadripper or Xeon. Personally I would lean more towards Threadripper. There are only a handful of motherboards but get one with enough pcie slots for all envisioned future expansion.
RAM speed is not going to be as much of an issue as capacity for your use case.
Cooling, make sure you have a large enough case with ample fans and ensure you understand the airflow path. There aren't many threadripper coolers out there but they all do a good job.
For storage you'll want a mix of NVME (for datasets you load frequently) and HDD for bulk capacity.
q-milk@reddit
I just built my second one, after learning from the first build.
You will move, load and store a lot of data.
About storage: After you pick an AM5 make sure the board has the PCIe 5.0 from the CPU directly to the M.2 slot, not through the bridge. No reason to not get the X870 chipset (for example GIGABYTE X870E AORUS XTREME). This double the speed of the M.2 to 16GB/s.
HobbyAnon@reddit
If you plan to run multiple GPUs in the future, keep in mind that most, if not, all consumer-grade motherboards can't run two GPUs at x16 at the same time. It's usually a combination of x8/x8, x8/x4 or x16/x4 (the other one is connected to the chipset like the ASUS B650E-E). Read the manual of the motherboard before you decide - some M.2 slots are shared with specific PCIe lanes which can downgrade the PCIe lane to x8 or x4 when both are occupied. Also, consider the slot spacing of the PCIe of the motherboard, the size of GPU and the case of the PC.
As for the RAM, I find 64 GB a bit small when running T2V/I2V models but it's mostly enough. 128 GB would probably fair better. As for the storage, AI models can fill it real fast - check the sizes of the models you would use. Gen 4 SSDs are more than enough.
PSU, in my experience, running 3 RTX 3090 on HX1500i was a bit unstable, even running just 2, I get occasional crashes when the moment inference starts but mostly stable when it's running. 2 RTX 4090 should theoretically run fine on 1500W PSU and probably better if it's power limited. You will need to buy additional compatible 12VPWR cable though.
iamapizza@reddit
So I assume you'll be running pytorch/cuda-based type workloads, or something similar?
Try to get as much VRAM as you can, so the 4090 sounds good, 5090 even better but the prices right now are mad.
If you go for 64 GB RAM try to get 2 x 32 GB. I think 64-96 might be a good range.
But I feel the storage will be more important especially if you start swapping to disk if your datasets are large, so a good speed 2 TB NVME is better long term vs 1 TB.
When you search for the motherboard, have a look at its manual, some motherboards will say they disable certain slots or reduce bandwidth if you plug in an NVME in certain slots. As long as it doesn't say that you are good, or as long as the limitation it introduces is not relevant to you then you are good.
For stability try to run Ubuntu/Mint as Linux is much better suited for long term stability and uptime with fewer interruptions.
Separate-Initial-977@reddit
I'm building a pc for similar smaller use case should I go with 2 5060 ti 16 gb vram or 1 5070 ti 16 gb vram ? 2x5060ti=1200$ 5070 ti =1100$
Conscious-Analyst662@reddit
5070 Ti. Two 5060 Ti cards only makes sense if you rlly need multi-GPU throughput. For most people, 2×16 GB is not a substitute for one stronger card, because VRAM doesn’t combine the way people hope. It’s more complexity for less simplicity and usually worse single-job performance.
VirtualArmsDealer@reddit
Don't worry about a 5090. Try to hit up the guy in China that makes 48Gb 4090. Also there is always the option of 96Gb h100 if money isn't a problem. Maximise vram where you can. 9800x3d and Samsung 980 pro are great.
prank_mark@reddit
Right now?
I'd get a 270K Plus with 64 or 128 GB of DDR5 RAM
If you're planning on running multiple GPUs, I'd get a Threadripper. But considering the cost, I'd only get that if you are sure you'll be getting more GPUs, and you're sure this will happen within a year or two. If you're not sure, get the 270K Plus for now and swap the motherboard and CPU later to upgrade to Threadripper.
YELLOW-n1ga@reddit
PCPartPicker Part List
YELLOW-n1ga@reddit
I couldn't find ram for your use case on pcpartpicker, you will need to find server ram with ecc buffer, which means no corruption during Long video processing, you do not want that. Preferable 4 sticks instead of 2 for max bandwidth. So buy 2 32gb/16gb kits. Ram capacity its up to you and how deep of a pocket you have. Ram speed isn't much of a factor so no need to go over 6000mhz
YELLOW-n1ga@reddit
PCPartPicker Part List
RationalDialog@reddit
Obvious solution would be threadripper with tons of RAM but it will be costly, especially now. RAM is costly, ssds are costly.
I don't know how big your datasets are but 64 gb RAM might not be enough. before you build the model you will need to do some type of clean-ups and selection I assume. that will use system level RAM.
Storage yes nvme and don't get crappy cheap ones. you want one with dram cache. believe me. ideally you get 2. one for the OS (can be cheap or even sata drive) and the fast one for the actual data so that OS background stuff don't interfere with data loading.
scalability is mostly about the motherboard (eg available slots and pcie lanes) and the PSU. There IDK if a ATX psu even exists that can run a threadripper and more than 1 5090. For sure it will be loud as hell, the entire machine under load.
XxNaRuToBlAzEiTxX@reddit
If you are planning on running multiple gpus in the future you might want to go for a threadripper build. It will depend on your use case, but they will have much higher PCIE bandwidth which you might need. I could be wrong, but I don’t believe there are any motherboards for the regular desktop cpus that can give 2 pcie 5 slots with 16 lanes for each
RumbleTheCassette@reddit
Yeah I was gonna say the same thing. The extra PCIe lanes and RAM support in general would be valuable.
Revolutionary_Ad7262@reddit
CPU choice should be either the best of the "normal" CPUs (like 9950x) or threadripper.
Normal CPUs are cheaper, faster and works well with a single 5090. Thread-ripper may shine, if you often utilize a full parallelism on CPU (where more cores -> better) or you want to have more PCI lanes
Same with motherboard
Impossible to answer without knowing your exact workflow
Just don't buy any leds, expensive cases and motherboard with a lot of feature, which you don't use
Buy NVM, everyone needs SSD. Later on you can buy more disks. it is not a decision, which needs to be made right now
NefariousSINNER@reddit
RTX 5090 24GB? What?
Alternative_Art2984@reddit (OP)
Mistake it's rtx 4090 24gb