Well they would require an awkward ROP/MC configuration to end up with 64 ROPs/ 2048b bus divided between six GPCs so that wouldn't really make sense but assuming that is a non-issue and they would have gone for 96/3072 it would provide 50% higher geometry throughput and a paltry ~15% increase in shader throughput thus shifting the balance significantly towards geometry, something that appears to be a major bottleneck for Vega
Yeah so we have two problem areas, the geometry throughput and bandwidth issues. bandwidth issues are up in the air, possible theory is increased latency of the longer pipelines, which might be able to be fixed by updated drivers, if its something like that.