GPU Scaling

The current deployment of GPUs on the UK Grid effec:vely relies upon containerising workloads that are then scheduled in a batch-like way. For users this almost certainly means that they must first gain interac:ve access to a GPU to develop, test, train, and debug their applica:on, before then moving to a different environment on the Grid. Such Grid-hosted GPUs are oMen significantly more powerful than local consumer GPUs but that, by itself, may not be sufficient incen:ve. In addi:on, the lack of standards for GPUs, mean that there are oMen addi:onal barriers in migra:ng work to other machines. So the work we propose is about streamlining the development environment, removing the hurdle of containerisa:on, and enabling GPU usage in an HTCcontext.