TL;DR My primary use case right now is image and video generation. Is it worth it to grab a 5090 prebuilt to move the parts between my homelab machines for generation workflows? (currently on an AMD R9700)
My my Linux desktop has an AMD R9700 (32GB) and I also have an older W6800 (32GB) on another machine. My primary use for them is Comfy-ui. I have bought zero RAM/hardware this past year due to these insane prices. I’ve got DDR5 SODIMM adapters and two pairs of mismatched RAM in my primary machine so I can have 95GB.
The ROCm drivers/libraries have gotten better, but AMD still falls behind and Comfy-ui sometimes works for a few days and then totally stalls out. Usually it just crashes the container, and sometimes it takes out my entire Xsession (xlibre/gentoo). WAN2.2, with most of my workflows, takes anywhere from 1~3 hours for ~10 seconds of video with decent workflows (the ones that hold a good amount of detail, even before upscaling).
I’ve held off considering buying an nvidia card. I have a 3080-Ti in my Win10/gaming box, but I game almost exclusively on my jailbroken PS4 and PS5 (air gaped, separate network) and am .. honestly not even interested in PC gaming anymore. I’ve been meaning to change out nvme drives and get Linux on it, just to see how much faster and/or stable an older nvidia gen card is for the stuff I’m doing, to see if it might be worth getting a 5090.
I haven’t been following 5090 prices until today and dear god! They’re getting near RTX 6000 prices from a few months ago .. and RTX 6000s are now $12k ~ $15k?! They were like $8 ~ $10k earlier this year right? I didn’t just imagine that?
But I noticed there are a ton of 5090 prebuilts for the cost of .. the card! Some of these integrator must still have stock from the previous prices and know they can’t push inventory if they start gouging. I bet a lot of people are buying these and piecing them out on eBay, but I’m honestly considering buying one just to have a 5090. Some of these have 32GB of normal RAM and 1~2TB SSDs for under $7k! That’s .. not bad right now.
There are some DDR4 servers from 2019 used that have a lot of DDR4 and are around $3k. Would it be worth it to put a 3080 + 5090 in such a rig to speed up local image and video generation? I’d like to get into some LoRA training too. I don’t do any local LLMs (use work credits and my personal openrouter account for coding).
If you are able to get a prebuild including card cheaper as full deal it can be a good way to save some money.
Only I’ll wont tell you how should spend it, but 7K is still a terrible deal in my book for a prebuild with your specs.
I know current market is ffed and a 5090 is 5k alone.
So if you need one currently, take advantage of whatever you can save to save a bit.
DO pay extra attention what kind of 5090 you’ll get in the prebuild.
Usually big system interrogators like Dell, HP, Lenovo have custom cost optimized versions of the GPU’s that’ll bundle with the prebuild.
And boy there are some stinkers sometimes they’ll make.
Not all of those are bad bit be aware you might get a blower design gpu, or really small downsized one with barley sufficient cooling. or just a retail equivalent but all black with no extras as rgb so they save a bit.
So check what reviews say about what is bundled by which OEM.
The smaller (boutique) system integrators Maingear, Cyberpower, etc will usually have retail cards in prebuilds.
But it might be a surprise what brand and model as they have to source left and right whatever they can get.
DDR4 based older servers are a good way to get access to large ram volumes for somewhat decent prices.
But do pay attention to any Xeon v3 and v4 servers, as those are mostly a waste of money.
A Xeon v4 is a based on broadwell, a CPU architecture more than 10 years old.
You’ll get lots of cores, but most LLM (I think) and home server stuff benefit from way faster single core performance.
A full explanation what to go for will be multiple paragraphs long as Xeon naming is a clusterF.
Intel has changed the naming conventions like 4 or 5 times in 6 years.
So you’ll need a literal decoder ring or wiki pages to check what you’ll get.
Before giving yourself a migraine on Intel label shopping check if any AMD epyc servers are available.
The Epyc 7003 (3th gen) was multiple years ahead on what Intel had to offer same period
Like a 16 core 7003 cpu beats a Xeon 32 core (same build year) in single and multi core hands down.
It’s the single fastest DDR4 platform available only beaten by Threadripper 5000, which is currently unoptainium.
So If you see any epyc 7003 server up for grabs go for it.
But be careful as 2dn gen Epyc 7002 is not bad but get’s beaten by Intel xeon newer scalable cpu’s.
So the picking order goed Epyc 7003, Xeon scalable 3th gen, Xeon scalable 2nd gen, Epyc 7002, Xeon scalable.
It it’s a Xeon with v3, v4, v5, v6 in it’s name not worth it.
Focus on the cpu’s with higher clock speeds on Intel’s side, AMD is all fast.
Intel has some serieus segmenting in it’s line up, usually only gold and platinum have clocks coming close to the 3Ghz.
Double check the Intel wiki what a CPU max all core load is as a 64 core 3Ghz gold cpu wil run full load at 2Ghz or while a platinum can do all core 2,6Ghz.
Also have a look at ebay for used server SSD’s there is still some decent priced stuff to be found.
Sometimes as low as 60-80 dollar a TB is you go for a pcie gen 3 SSD.
But even some gen 4 still are findable for below 100 dollar a TB
Yep, I’ve seen plenty of GN reviews where a prebuilt has a generic Dell/HP nvidia card, sometimes with insufficient cooling and sometimes where it’s fine but just looks basic.
The cheapest ones on Newegg (~$6.5k) all appear to have retail cards (MSI on one, the Thermaltake one doesn’t list the card; might have to dig into that one if I go that route).
Yea I was looking at some of the 2019 Macs since you can get adapters for PCI-E and there is an ebuild for a kernel for running Gentoo on the T2 chips. But the EYPCs you mentioned seem reasonable and available.
Huh, I guess that would explain the current 5090 pricing. I remember something similar with the 4090 right before the 50xx releases. The trouble with the whole “just wait” is it could just as easily just get way worse. I saw the numbers of the wall when I took the plunge and upgraded my storage array in December:
There’s still no way I could by 4x30TB today at those prices.
GPU prices were getting closer to reasonable at the tail end of 2025 before the RAM-pocalypse.
I’d rather not buy anything in this market, and maybe I will just hold off for a few more months. $7k is a lot to spend right now, and 8 months ago a little more than that would get an RTX 6000. It’s honestly a crapshot at this point to guess if prices will start to look more reasonable by 2028, or if there will even be consumer hardware left to buy.
It’s not the single core performance. It’s the RAM channels. Those C602/C612 chipsets are all quad channel and max out at 2133/2666Mhz respectfully (I overclocked some to 2933 in V4).
Nah the real specs you are looking for are 8-channel memory CPUs, or higher. Think Epyc (even the original Epyc is 8-channels, and can overclock to 2933 Mhz). I measured 110 GB/s ram with an 7551, and a Rome CPU in the same Mobo and ram clocked in at 128 GB/s. That was single socket - dual socket Rome I’ve seen people post 205+ GB/s memory bandwidth on reddit.
CPUs are dirt cheap ($20-$50). But it’s the Mobo that will break the bank. Even the old original AsRock Rack EPYC8D (7 x 16lanes of Gen3) commands $800+ on ebay. ROME8D that is Gen4, $1000.
Ps: the original Gen1 Epyc are all unlocked. You can ram up the GHz and/or TDP wattage accordingly, as I did with my 7551 to 400W full turbo to 3.4 GHz on all cores.
That’s all LLM inference cares about, memory speed not core speed.
However, the OP is asking about image and video generation. So the point is mute because no one encodes on CPU for those.
Warranty for the parts could be a problem if any of the shucked components fail as warranty would be handled by the system integrator, not the card manufacturer.
For video and image generation you really cant beat a 5090 right now. I have no complaints on that front. Diffusion models are small enough you dont need more than 32gb yet. Less is tricky though even with quantization.
I’ve been meaning to change out nvme drives and get Linux on it, just to see how much faster and/or stable an older nvidia gen card is for the stuff I’m doing, to see if it might be worth getting a 5090.