The scheduler works with the ROPs directly, and in the past the ROPs could not do out of order and had to be scheduled in clusters. One of the big reasons that AMD chips did not scale well when they got larger (vega/fiji) was the way the rops worked. It required too much latency so they were bad at high framerates. They had it detailed in the RDNA2 white paper. There was coverage but all I am finding now is about the rx6800 and not the rdna2 reveal.