I want to use linux as my main OS, and windows for gaming & 3D CAD only.
Recent higher-end GPUs are really huge (~3slots), and I really don’t want to do a custom loop.
Is a L1T KVM a good idea for this (dedicated linux PC + windows PC in parallel) use case?
Long version
Background
So I was really fascinated by Wendell’s video about GPU passthrough and lookingglass, and wanted to do my own.
While I bought a mobo and CPU for this purpose, COVID struck and GPU prices were atrocious.
Recently since prices are somewhat acceptable, I upgraded my GPU and realized how huge they are, while technically 2.8 slots, It realistically needed 4 slots for cooling!
Looking forward, I don’t think GPUs will get smaller. Which means if I want this level of performance and do dual GPU with passthrough, I’ll need to do custom watercooling - which I really don’t want to do.
I want to let my PC be maintenance free, except occasional dust filter cleaning.
My current ideas
So I think I’ve 3 options:
Swallow the pill, do a custom loop (at least for the GPU)
Buy an AM5/intel system with iGPU, pass the dGPU and use iGPU for host OS
Buy a L1T KVM(DP1.4), and buy/build a second SFF machine as primary linux PC (or secondary windows machine)
And I’m starting to think that a L1T KVM might be the most simple setup.
After all, I don’t have to worry about programs complaining that it’s in a VM.
I also have a NAS now, which means shared storage really isn’t a problem.
What I want to ask
Is anyone using a L1T KVM for this use case?
If yes, can you share your experience?
My current hardware
Monitor
4K 144Hz using DP
PC
mobo: MSI MEG X570 unify (PCIE 4.0 x8x8, 1st x8 slot has 3 slots before 2nd x8)
case: Fractal meshify S2 (7 slots, but the top slot on the mobo is unused, so 3+3 slots for PCIE x8)
CPU: 5900X (bought to use the 2nd CCD for the VM)
GPU: 2070 super (1.8 slots) → used 6800xt (2.8 slots)
I basically came to the same conclusion a while ago.
While Looking glass is neat, two GPUs are just too much especially as a GPU today can cost about as much as a whole new midrange system. A $800 Radeon XT can easily be paired with $800 rest of components. This means we are looking at $2400 for a dual GPU system versus $3200 for two identical systems. It’s “only” 33% more money for two systems.
Also, APUs are starting to become good. It is quite possible next generation you don’t even need to choose between a 12 core or an APU, though I will believe it when I see it. This eliminates the need for a second GPU.
So, if you have an older GPU laying around as a spare, by all means use it, but I find it hard to justify VFIO when you can just run two midrange rigs for similar overall performance and much less hassle. Just my 2 cents.
For a secondary/new primary system, recommend a Ryzen 7900 + 64GB RAM + 2TB SSD + B650 motherboard as a core. For Intel, Z690 boards are the only sane mid PC choice, then get a 13600k and go to town with same RAM and SSD.
I used a 5800X w/ 32GB RAM and a RTX 2070 passed through to a Windows VM with a RX 5500 XT used by the Fedora host.
Looking Glass is amazing software with no noticeable latency, especially at >60hz, however, I would recommend using two machines and a KVM. I was constantly running into little quirks and weird performance hiccups that were difficult to trace.
I now have a Windows desktop for games and a Linux laptop for productivity hooked up to a L1T KVM. Highly recommend.
Is your RTX pemanentely unsused by linux ?
i have an Amd GPU with an extra 3080TI, i Need it to play with Cuda and to drive one of my display for everyday task, but i cannot unload the kernel driver to pass it to the VM.
To me, a dedicated GPU for windows sound unresonable givent that 90% of the time i’m under linux, and not managing to pass it mean that it’s easier to move the GPU from one PC to the next when i need to play game unsupported in proton.
Thank you all for your comments! This is really an incredible community.
Since I live in a space-constrained area, I might even sell my current primary system!
Maybe I’ll make an mATX build with the ASUS AP201, which my GPU should fit into.
The Linux system might as well be my laptop!
May I ask for the latency for switching between those 2 systems? (especially with mixed refresh rates between PCs)
Also, what are your strategies for sharing data that are sensitive?
e.g. clipboard, password manager (i use bitwarden)
Noted!
I thought about proton too, I own a steam deck and it’s absolutely amazing!
Though I decided to ditch that idea, since I tend to play newer games, and it’s still hit or miss for latest (not so well known) titles.
Yes, when I was using this setup the NVIDIA card was blacklisted from the Linux kernel.
It’s fairly quick, almost instant. Both systems run the monitor at 144hz. This may largely depend on your monitor.
I don’t. Windows doesn’t see any sensitive information. It runs games and voice chat. Both Steam and Discord support QR code login via mobile app.
Edit: If you do end up going the KVM route be sure you pickup short, high quality, DisplayPort cables. I was seeing intermittent handshake, resolution, and refresh rate issues until buying recommended cables.
Yeah makes sense, but it’s good to hear about the ideal results!
If you have time, can you please check what it’s like switching between 60Hz and 144Hz PCs?
I know the answer will be “depends on the monitor”, but just for a point of reference.
Didn’t think about QR code login, should’ve thought about that.
Yup, definitely going to pickup some Club3D 1m cables! (it’s cheaper than L1T cables in japan, even before shipping)
We try to discourage this configuration, the iGPU uses system ram as GPU ram which causes bandwidth issues when you start to try to push high FPS and/or resolutions via Looking Glass.
If you are only doing high graphical loads in the VM (games) then your host GPU can be anything really, just get a workstation GPU that is only 1 slot wide.
Note too that LG allows the integration of your guest into your desktop workspace, where a KVM switch means you are limited to one or the other, not both at the same time. You also get clipboard sharing and guest audio via LG, so you wont have to mess around with extra cables/mixing, etc.
The bus isn’t the issue, the memory bandwidth is. LG has to copy ram from the guest GPU to system RAM, then from system RAM to the host GPU.
If the host GPU is also using system ram, we are now doing an in-place CPU copy from system ram to system ram. This is 50% slower then copying from ram to GPU as 50% of the time will be spent reading vs writing.
We do have methods to mitigate for this like using DMABUF but it’s not a solution, just helps.
Have not looked in a while on what’s available. but have you considered a GPU with an integrated water loop?
This brings the size of the card back down to a single slot, it might cost a bit more but in the long run might be a better option overall.
I personally run a custom loop so I just get the cheapest GPU I can get that someone reputable has made a full GPU water block for.
Edit: I think it’s a farce that the heat sinks are getting so large, they have a thermal problem and they are not facing it. And once the GPU gets heat-soaked your performance plummets anyway.
Edit2: Just something interesting, if you can keep it cool this is what a 3090 can do.
Yeah, but they’re either too damn expensive. My price range was under 300$, and now 400$ - which is why I ended up with a used card.
Used AIO GPUs are unobtainanium, and new ones are ridiculously priced.
I guess that’s the only way to have 2 dGPUs these days
But I really prefer air cooled cards, after all.
It is, though somehow I can’t think it’ll be better. I’d really love an “downclocked” lower TDP & smaller cooler but same die & memory version for the higher-end GPUs.
5-10% performance reduction, but <200W would be nice.
Can’t watch that, it was hidden. But I feel I know what you mean.
Depends how much you are willing to invest and look for. The Powercolor 6700 XT Fighter, for instance, is very small and nifty, as are most Nvidia 2-slot Founder Edition cards.
