Sharing GPU over network

Hi all,

I recently watched the video on progress for being able to share a GPU across virtual devices and I began wondering if there might be a time when we will be able to share a GPU not just across virtual devices but also across the network, especially as our local network connections get faster and faster.

i would imagine you would have a "resource server" so to speak on the network, and then all other client computers would be able to utilize the resource from that server.

Such things all ready exist and have done for quite a while. Look into "Thin Clients": https://en.wikipedia.org/wiki/Thin_client

The thing is user desktops are far behind in terms of features we are allowed have/reasonably pay for. The type of GPUs that allow this, and the rest of the parts for the servers that they would go in, typically run in the regions of thousands of dollars per part. So they are only feasible for entire company scale user loads, you would need hundreds or thousands of users to make a system like this viable cost wise compared to hundreds of individual PCs for each person.

Steam in home streaming works quite well for gaming over a LAN remotely at 1080p 60fps. You can game on a low end laptop that only has integrated graphics using a gaming PC on the same LAN to actually do the gaming computation and GPU work. The laptop would just become a thin client of sorts in this scenario.

In my case, I have an Unraid based headless server with a windows 10 VM with GPU pass through. I also have Steam Link boxes attached to various TVs around my house. My kids can "share" this server GPU and play games on the TV using this. No sound of loud fans to disrupt the experience.

I understand that Nvidia Shield can do this similarly although I don't own one. Nvidia also has a service that you can rent their gaming servers and play games on their servers over the internet using low end computers that lack a gaming GPU.

I've looked into playing over VPN using my steam streaming server over the internet, unfortunately my ISP does not provide enough upload speed to make this viable at this time.

So in a way, at least for gaming, it is possible to "Share" your GPU over a local LAN, and if lucky enough to have enough badwidth from your ISP, over the internet as well.

1 Like

I see, thanks for the replies! It's too bad that kind of technology isn't available in consumer products. Unraid will probably be the best option for me (or something similar when the time comes).

I believe this is kind of how mainframes and flight simulators worked back in the day with GL. the mainframes did all the math then piped the GL instructions to the workstations over a network. I may be wrong. Now the GPU's can do all the math and drawing on a single card. Are we just moving backwards to the old unix/SGI days?

Also check out No Machine and the NX protocol for thin clients you can pretty much make out of anything even raspberry pi's, the compression is insane. I used it on a gigabit ethernet network to play skyrim of my desktop on a crappy core2 thinkpad running Linux. for the most part aside from a few skips it was highly playable. I would not attempt this with any first person shooter games where the slightest bit of latency is death for everything else even watching videos it's very decent. The free version is lan only but you can get around that with something like the free zerotierone p2p vpn client.

I've been waiting for the SteamLink2 or whatever to do 4k@60fps, now that would be something.

1 Like

I think Shield TV can do 4k 60fps but requires a newish Nvidia card

No Machine looks great!

It works really well, Iv'e watched it evolve over the years, for gaming you want to match the resolution in the game to that of the thin client your using. Wireless performance may vary as expected, If you have a decent ethernet network you can pretty much almost build what the steam and shield devices are doing yourself. NO machine uses libraries on the client side for a speed up. I think the steam/shield do similar except piping some the raw DX or OpenGL to be executed by a local GPU. it is really similar to the old mainframes/workstations from 30 years ago.