There actually _could_ be an explanation for this if you are referring to the doom/vulkan benchmark. The RX 460 is only 1/4 the performance of the RX 470 there whereas you'd expect it to be about half. It could be legit if the slowdown is due to massive texture swapping because of insufficient amount of memory - and in this case the RX 460 could be slower due to only having pcie 3.0 x8 connection.
Okay that's a fair bit smaller than I thought.
Both the new cards have half the bus width of the previous generation products in your example. So there are significant increases in perf/GBps (architecture) and in perf/W (process+architecture).
After some additional polishing and improved illumination it looks like this (click on it for full resolution):A german madman polished down the die of a RX 470 (more here http://www.forum-3dcenter.org/vbulletin/showpost.php?p=11124532&postcount=10484)
36 CUs confirmed.
). But it was enough to count the CUs.
The delta color compression is better in Polaris right? TechReport shows DCC in P10 being almost as good as Maxwell which is pretty damn good
I just linked to it. You have to thank OC_burner from the 3DCenter forum (or 'Fritchens Fritz' on flickr, same guy). He does all the work. As I said, it's pretty amazing what quality he gets without resorting to professional tools. It's a hobbyist's effort. But he definitely gained experience. The newer pictures are way better than the early ones just a year ago. Btw., on flickr he also shows work in progress photos.
Most of the improvements are small interative tweaks. The CUs are still highly similar to GCN1 (Radeon 7970). Nvidia's Fermi -> Kepler -> Maxwell -> Pascal (P100) are each (on their own) a bigger change to compute units than the whole GCN1 -> GCN4 iteration. Maxwell -> Pascal (gaming) is mostly minor tweaks + shrink + more CUs + lot higher clocks (to already great Maxwell architecture).
