cross-posted from: https://lemmy.ml/post/10454803
Further if this technology is open-sourced; can it be extended for use cases beyond that(Dual Motherboards sharing Compute power with low latency for working on a single process?); I know such solutions probably exist for servers and enterprises but i am talking about amateurs who don’t have 10K lying around for specialty hardware: If possible this seems like a low cost solution to mess around with
RDP with low latency over thunderbolt? from the video it looks its new software intel has developed for windows, so its most likely proprietary. I was mostly thinking along the lines of using the technology to simulate S.L.I where half the frames are drawn by one pc and the other half by another
Also in the Article the data transfer speeds are in Mbps whereas in the video it is touted in Gbps
With GStreamer you can build a pipeline you like, you don’t need to use RDP, you can send uncompressed frames plain over network like in the video. I’m not an expert on graphics processing. SLI or NVLink are (I think) proprietary parallel processing interconnects. But NVidia didn’t invent parallel processing. I’m sure there are other solutions available. Though, I somehow doubt those will help you because they’re generally tailored to other (HPC/datacenter/simulation) purposes and not for gaming. And I think they use something like Infiniband for that and not thunderbolt.
With the speed, mind the first article is 5 years old. And I’m not sure how the hardware in the second one compares to what Linus uses or if it’s even the same generation of Thunderbolt. It’s probably gotten way faster since. I can’t try because only 1 device I own supports thunderbolt at all.
I think transferring files over thunderbolt networking or low latency video is nothing new. It can be easily replicated. And setting up 2 gstreamer pipelines is just two (lengthy) commands. Replicating NVlink is another thing, though. We probably need an expert on graphics drivers to tell if that already exists or how difficult that would be to implement. Most people will probably just fit 2 graphics cards into one computer or buy one faster GPU because that is both cheaper and way faster than connecting them in 2 separate computers with added latency.
(MPI would be an example of an open standard to do parallel computing with arbitrary interconnects.)