From
Anandtech's review:
So basically, RTG's engineers are claiming they
could have increased the number of compute engines (and thus the number of triangles/clock) if they wanted, but they just didn't care. This reinforces my belief that they are basically giving up on gaming. If they were taking gaming seriously, they would have increased the number of compute engines to 8 and the number of ROPs to 128. They had plenty of room to do this while still fitting within the reticle limit, and it would have prevented disgraceful results like this one:
Instead, we get a card that basically retains almost all of the potential bottlenecks of Fiji. In a handful of cases, the drivers can try to use tricks to reduce the amount of geometry rendered, but when these optimizations are not done or when the geometry actually does have to be drawn, then Vega is no better than Fiji.
I don't expect anything worthwhile in gaming for Navi, either. They aren't going to bother to make two GPUs act transparently as one - they'll expect the developers to do it with explicit multi-adapter in DX12. The developers won't; instead, they will continue to write easier-to-code DX11 titles that Nvidia's driver team will optimize to the point that they are as fast as DX12. Navi will still be limited to four compute engines and 64 ROPs because RTG simply doesn't care to do any better.