Handling pipeline submissions across many compute units

A method comprising:

• distributing a plurality of thread groups across an array of multiprocessors, the array of multiprocessors coupled to an interconnect fabric comprising one or more fabric switches to couple the array of multiprocessors to a plurality of memory devices;
• assigning a multiprocessor of the array of multiprocessors to process a first thread group comprising a first plurality of threads;
• executing instructions of a first thread sub-group and instructions of a second thread sub-group via circuitry of the multiprocessor, the first thread sub-group and the second thread sub-group formed based on the first thread group and respectively include a plurality of threads, the second thread sub-group having a data dependency on the first thread sub-group;
• and
• causing threads of the second thread sub-group to sleep until threads of the first thread sub-group have satisfied the data dependency via circuitry of the multiprocessor.

A data processing system comprising:

• a memory device;
• and
• an accelerator device comprising a processing cluster including an array of multiprocessors coupled to an interconnect fabric, scheduling circuitry to distribute a plurality of thread groups across the array of multiprocessor, and a multiprocessor of the array of multiprocessors to be assigned to process a first thread group comprising a first plurality of threads, the multiprocessor comprising circuitry configured to:
• execute a first thread sub-group and a second thread sub-group, the first thread sub-group and the second thread sub-group formed based on the first thread group and respectively include a plurality of threads, the second thread sub-group having a data dependency on the first thread sub-group;
• and
• cause threads of the second thread sub-group to sleep until threads of the first thread sub-group have satisfied the data dependency.