Most “GPU”s found in datacentres don’t have any graphic connectors on the back of them. They do all their communication via PCIe.
In some cases they may have a connectors for high speed communication fabric between many GPUs. NVLink is an example of this [0].
Outside of the consumer and workstation space GPUs really don’t look like anything you or I would recognise as a GPU anymore, and with APIs like CUDA, don’t look like GPUs to software either.
Really all we’re missing is open source drivers. But I wouldn’t hold your breath.
In some cases they may have a connectors for high speed communication fabric between many GPUs. NVLink is an example of this [0].
Outside of the consumer and workstation space GPUs really don’t look like anything you or I would recognise as a GPU anymore, and with APIs like CUDA, don’t look like GPUs to software either.
Really all we’re missing is open source drivers. But I wouldn’t hold your breath.
[0] https://www.nvidia.com/en-us/data-center/nvlink/