OK, so there's a bit of a debate about it recently and a few tech YouTubers have also given their opinion, so it's only relevant that some random nobody with too many posts on a tech-forum also gives theirs.
I'll start off with what GCN cards AMD has released since 2012.
R9 390X - Hawaii XT 2816 Stream processors based on GCN 1.1 R9 390 - Hawaii PRO 2536 Stream processors based on GCN 1.1 R9 380 - Tonga PRO 1792 Stream processors based on GCN 1.2 R7 370 - Pitcairn PRO 1024 Stream processors based on GCN 1.0 R7 360 - Bonaire PRO 768 Stream processors based on GCN 1.1
R9 290X - Hawaii XT 2816 Stream processors based on GCN 1.1 R9 290 - Hawaii PRO 2536 Stream processors based on GCN 1.1 R9 285 - Tonga PRO 1792 Stream processors based on GCN 1.2 R9 280X - Tahiti XT 2048 Stream processors based on GCN 1.0 R9 280 - Tahiti PRO 1792 Stream processors based on GCN 1.0 R9 270X - Pitcairn XT 1280 Stream processors based on GCN 1.0 R9 270 - Pitcairn XT 1280 Stream processors based on GCN 1.0 R7 265 - Pitcairn PRO 1024 Stream processors based on GCN 1.0 R7 260X - Bonaire XT 896 Stream processors based on GCN 1.1 R7 260 - Bonaire PRO 768 Stream processors based on GCN 1.1
7970 - Tahiti XT 2048 Stream processors based on GCN 1.0 7950 - Tahiti PRO 1792 Stream processors based on GCN 1.0 7870 - Pitcairn XT 1280 Stream processors based on GCN 1.0 7850 - Pitcairn PRO 1024 Stream processors based on GCN 1.0 7790 - Bonaire XT 896 Stream processors based on GCN 1.1 7770 - Cape Verde XT 640 Stream processors based on GCN 1.0 7750 - Cape Verde PRO 512 Stream processors based on GCN 1.0
Versions of GCN and improvements
GCN 1.0 - Initial release with the 7000 series in 2012, uses a SIMD (Single Instruction Multiple Data) arcitecture instead of the old VLIW design used in previous arcitectures like the 5000 and 6000 series. Each GCN compute unit totals 64 stream processors. AMD launched it as a long term building block for future GPUs, which is why they still use it today.
GCN 1.1 - GCN 1.1 was a minor iteration that added feature support for TressFX and TrueAudio, and improved PowerTune support. Most of the improvements were compute and HSA focused for the APU side of things.
GCN 1.2 - GCN 1.2 is a release we have only seen in Tonga and Fiji so far, the big focus with the release of the R9 285 was power efficiency improvements, AMD improved their instruction scheduling and added newer instruction support, the additional 16 bit Floating point & Integer instructions means that less power can be used on applications that do not need the mathematical accuracy of 32 bit instructions. The new Fiji GPU has a reworked memory architecture so it works with the HBM memory.
Improvements in manufacturing
So a term thrown around a lot is "refinements", unfortunately no-one has been too specific about the refinements but there are board level improvements and the process used has seen improvements.
Both AMD and Nvidia have used the same TSMC 28nm manufacturing process since 2012, over time as the process matures the ability to produce big chips consistently improves, and performance consistency also improves. We get less leaky chips. Which means we can have higher stock clocks, better overclocking and the better chips achieve this with the same or less voltage than before, which leads to a lower average power draw. This is very similar to Intel's Haswell Refresh/Devil's Canyons (note how people are fine with this being called refresh and not rebrand?).
The new cards have also received higher clocked VRAM, from 1250 to 1500MHz.
AMD have generally been more generous than nvidia when it comes to VRAM in the past. This is still true for the most part, the Hawaii based products now offer 8GB of VRAM, AMD were saying a lot about Crossfire technology and have improved crossfire support in their new drivers as well, AMD would like people to pair multiple R9 300 series cards to play games at 4K and above resolutions and they've given us the framebuffer to achieve this, reviews have demonstrated that above 1080/1440p the R9 390 and the 970 begin to make more of a gap in the 390s favour and some games will cap out the 970s effective 4GB VRAM. They also offer a 4GB version of the R9 380 at a similar price to nvidias 960 4GB.
Here I'll list the MSRP launch price of the products from the 200 and 300 series and the 900 series nvidia cards, a lot of people are saying you might as well pick up the 290X as it's very close to the 390X and is cheaper, which is true. You also need to consider that the 7970 was cheaper than the R9 280X until it went out of stock... very quickly. So if you want to buy you may want to buy now, as this will not hold true for long. Note I am only listing the 200 series cards that have a 300 series equivalent.