Originally Posted by Mahigan
There are two models the C4 and the C7 announced so far. The C4 and the C7. The C7 has a clockspeed of 1266MHz and is known as the RX 480. In 3D Mark.. its performance is similar to a Radeon Fury though folks are using its TFlops figures in order to place it as an R9 390 replacement (not understanding that AMD have optimized their GCN cores and caches in order to boost efficiency in that metric not unlike the nVIDIA Kepler to Maxwell changes
).. This leaves the door open for an RX 480x which should have 2560 cores compared to the RX 480s 2304 cores. The RX 480 and 480x both having 64 ROPs. I am thinking that we are going to see Radeon FuryX performance out of the RX 480x.
36 CUs leaves us with 144 Texture Mapping Units for the RX 480 and 40CUs leaves us with 160 Texture Mapping Units for the RX 480x. The RX 480x will probably go for $249-$279. This leaves the door open for a $349 and a $399 set of products which will be based on small Vega (RX 490 series) as well as the premium big Vega GPUs which are likely to be priced at $549 and $650 respectively (RX Fury-like lineup).
The RX 490 series will be going head to head with the GTX 1070 and 1080 offerings from nVIDIA while the big Vega GPUs will likely shoot for GTX 1080 Ti and Titan territory.
So the performance per CU is up big time. We will have to wait for the NDA lift in order to ascertain by just how much (I figure that a good chunk of that 30-40% IPC boost is from the changes to the CUs organizational structure). That being said... the RX 480 series are limited by their use of GDDR5. This could be particularly true for the RX 480x.
Then the full Polaris 10 die isnt out yet? if so then RX 480 has 8GB and 2304SP and 1,26Ghz, RX 480X has 8 GB and 2560SP , then what would be the CU/Shader count for the 490(x) to match a 1070/1080? 3072-3584?with 16GB GDDR5@384bit?
Originally Posted by sugarhell
The biggest problem is that they compare the theoretical TFlops to get a comparison. All the previous GCN cards have really high peak Tflops performance but on games they can't peak so high.
I suspect that polaris can hit the peak performance much easier than fiji.
Also i suspect that polaris is so small because it's mostly a gaming card.
well people should compare to a 380(x)
RX 480 with 1,26GHz and 2304SP has 5.8TFlops, R9 380 with 1792SP and 970MHz has 3,46TFLops there is a 58.5% difference
RX 480X with 1.26GHz and 2560SP has 6,48TFlops, R9 380X with 2048SP and 970MHz has 3,97TFlops there is a 63.22% difference
maybe if there were a better GPU people wouldnt be comparing them
Originally Posted by CataclysmZA
So, looking at things right now, and given what we know about the RX 480, is Ellesmere a die shrunk version of Hawaii? I think it could be.
The R9 390 is 5.8 TFLOPs, while the RX 480 is about 5 TFLOPs or faster, depending on who you talk to. 36 compute units and 64 ROPs sounds like a cut-down Hawaii chip, and 40 CUs with a die shrink and a clock speed boost would easily yield GTX 980 Ti-like performance (if not beat it soundly and eat its children for breakfast).
If the rumored RX 490 is $299, NVIDIA might have to drop the price of the GTX 1070 early into the card's life cycle, but not too early to annoy their partners and customers.
Polaris is more than a nodeshrink. it isnt like Pascal using similar pipeline as Maxwell, GCN4 like mahigan said adds more cache L2,improved geometry processor(he claims that AMD added some kind of back face culling like Nvidia gpus has), better GCP and keep/improved Memory compression From GCN3Edited by PontiacGTX - 6/2/16 at 9:19am