Abstract: A data processing system comprising a plurality of processing nodes, each comprising at least one memory configured to store an array of data items, wherein each of the plurality of processing nodes is configured to execute compute instructions during a compute phase and following a precompiled synchronisation barrier, enter at least one exchange phase. During the at least one exchange phase, a series of collective operations are carried out. Each processing node is configured to perform a reduce scatter collective in at least one first dimension. Using the results of the reduce scatter collective, each processing node performs an allreduce in a second dimension. The processing nodes then perform an all-gather collective in the at least one first dimension using the results of the allreduce.

Abstract: A system and method for providing a set of data transfer instructions for converting one or more tensors between two different layouts. A first layout is used for storage of the data in host memory. A second layout is used for storage of the data in external memory accessible to a subsystem. The subsystem acts as a work accelerator to the host, and reads the external memory and processes the data read from the external memory. The first layout may be a logical representation of the tensor. The second layout is optimised for transfer to and processing by the subsystem. The data transfer instructions for converting between the two layouts are generated in dependence upon an analysis of the instructions to be executed by the subsystem.

Abstract: A data processing system comprising a plurality of processing nodes, each comprising at least one memory configured to store an array of data items, wherein each of the plurality of processing nodes is configured to execute compute instructions during a compute phase and following a precompiled synchronisation barrier, enter at least one exchange phase. During the at least one exchange phase, a series of collective operations are carried out. Each processing node is configured to perform a reduce scatter collective in at least one first dimension. Using the results of the reduce scatter collective, each processing node performs an allreduce in a second dimension. The processing nodes then perform an all-gather collective in the at least one first dimension using the results of the allreduce.