GPUs become requestable resources in cluster schedulers (2010s)
As GPU accelerators spread through academic clusters, workload managers added generic resource tracking so a batch script could ask for one accelerator on one node. Before that, labs reserved whole GPU machines or managed access by hand.
GPU jobs became a line in a submission script instead of a separate arrangement with whoever ran the hardware.
Shared clusters could serve many labs running accelerator code, which is the arrangement most machine-learning research still depends on.
The pattern repeats here: a new kind of processor gets a resource definition, and adoption follows the scheduler rather than the vendor.
