We're already there. That's what a lot of people are using DPU's are for.

An example, This is storage instead of GPU's, but as the SSD's were PCIe NVMe, it's pretty nearly the same concept: https://www.servethehome.com/zfs-without-a-server-using-the-...

PCI-e Networks and CXL are the future of many platforms... like ISA backplanes.

	▲	0manrho 4 days ago \| parent [-]
		Yep, I have a lot of experience with CXL devices and networked PCIe/NVMe (over Eth/IB) Fabrics and deploying "Headless"/"Micro-Head" compute units which are essentially just a pair of DPU's on a PCIe multiplexer (basically just a bunch of PCIe slots tied to a PCIe Switch or two). That said my experience in this field is more with storage than GPU compute, but I have done some limited hacking about in the GPGPU space with that tech as well. Really fascinating stuff (and often hard to keep up with and making sure every part in the chain supports the features you want to leverage, not to mention going down the PCIe root topology rabbit hole and dealing with latency/trace-length/SnR issues with retimers vs muxers vs etc etc etc). It's still a nascent field that's very expensive to play in, but I agree it's the future of at least part of the data infrastructure field. Really looking forward to finally getting my hands on CXL3.x stuff (outside of a demo environment.)