AMD are stuck with a very uncompetitive Polaris architecture (atleast in its current implementation) as we can see this from the pricing compression of the P10 product stack.
I agree.
Ideally if P10 had been able to hit clocks of 1400 Mhz for Rx 480 while staying below 150w and 1300 Mhz for Rx 470 while staying around 120w AMD could have priced Rx 470 4GB at USD 199 and 8GB at USD 229.
Um, so you're saying if Polaris 10 were a totally different, better-designed GPU AMD could have gotten more for it? OK.
The Rx 480 4GB could have launched at USD 239 and the 8GB at USD 269. Selling a 232 sq mm die for < USD 200 (even if its salvaged dies) at this early stage of the node's life is painful from a cost and profitability point of view.
It's probably a bit painful, but the real problem is that the reference boards are over-engineered with a lot of power phases, traces to support 256-bit memory bus, 8GB of RAM, etc. The whole bill of materials cost is quite high but the performance puts a really low price ceiling on the products.
Right now AMD will sell every Rx 480/Rx 470/Rx 460 they make.
This statement is not very helpful. I would assume that AMD has a reasonable handle on demand and won't overproduce chips.
But they desperately need a second revision to improve their perf and perf/watt so that when supply catches up with demand and when Nvidia has enough supply to flood the market with cheaper GP106 products and GP107 products AMD are not left relying on price cuts to move product.
Second revision? At this point they would need an entirely new architecture because this is pretty much it. The process that they're using has been in mass production for quite some time, so don't expect miracles on that front. And, unless you think AMD is going to rework the entire physical design of the product to squeeze out more perf/watt, this is pretty much it, what you see is what you get with Polaris.
Anyway, supply of GP106 seems to be quite good, they seem to come in stock much more frequently than RX 480 cards do, and they are available from a much broader set of AIB partners than the RX 480s are. I do agree that as 1060 supply gets better, aside from sales to miners, RX 480 sales will not hold up well.
Polaris has been a disappointing launch partly because of the current state of GF 14LPP. AMD needs a way out of the WSA and GF need to be told they need to have a competitive node and robust implementation if they want AMD wafer orders.
Don't blame GloFo for this. NVIDIA spends a lot more in R&D on its GPU architectures than AMD does and this is the result. These projects also began years ago, so what you're seeing now is really reflective of the relative R&D spending that we saw in the ~2013 time-frame. Since then, NVIDIA has ramped up spending while AMD has cut spending (even as the latter spreads its efforts across a wider range of products while the former has actually put an end to entire product lines such as Icera baseband):