Originally Posted by Yttrium
I dont understand how you got to the 980GB/s number for vega.
And your own numbers for 'effective bandwidth' are wrong (effective bandwidth is rather disingenous for enthousiast but I'll get to that ) your own link states that the 1080 has 1.7 greater 'effective bandwidth' compared to a 980 without any compression, something the 980 has but is not taking into account by Nvidia in this slide. the actual benefit from compression is 1.2 as stated in the slide and as stated by the reviewer. "NVIDIA pegs the effective increase in memory bandwidth from delta color compression alone at 20%. The difference is of course per-game, as the effectiveness of the tech depends on how well a game sticks to patterns"
which means the 1080's 380GB/s has 380*1.2 = 456GB/s
Then you talk about the 1080 Ti and a factor of 2.5 for compression, If you could elaborate more on this that would be great.
PS, on the same page it mentions bandwith over GPU power is getting lower over generations, how would it be bad if Vega follows that trend?
Why don't you read the article again about delta color compression...
The 2.5X of Pascal is based on no compression. It's 1.2x the benefit of Maxwell (GTX 980/GTX 980 Ti).
The GTX 1080 has 256-bit memory interface operating at a 10Gbps = 320GB/sec for reference cards. 320GB/s x 2.5 = 800 GB/s when there is DCC involved (such as graphics).
The GTX 1080 has 1.4X the raw bandwidth (i.e. no compression) from going to 10GBps instead of 7 Gbps GDDR5x.
The GTX 1080 Ti uses 11Gbps GDDR5X on a 352-bit memory bus = 484 GB/s , with 2.5X color compression = 1,210GB/s
To obtain 1,210GB/s with VEGA assuming Polaris level 1.3X DCC , you need ~ 930GB/s.
This DCC doesn't apply to compute workloads.
If you recall Tahiti HD 7970 / R9 280X transitioned from 384-bit to 256-bit on Tonga (R9 285) by using delta color compression.
Originally Posted by CynicalUnicorn
Well... Vega and Fiji both have 4096 cores. A Fury X runs at 1GHz while Vega hits 1.2GHz boost. Polaris is about 15% faster per core per clock than 28nm GCN. TPU has overall benchmarks for the reference 1080Ti and some older cards here: https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_1080_Ti/30.html
A Fury X is 60% as fast as a 1080Ti. Vega therefore should be (0.6)(1.2)(1.15) = 83% as fast, putting it right around a reference 1080.
I don't think Vega will be competing with Nvidia's high-end, unfortunately, but it should do as well as or maybe a bit better than GP104. I think AMD's biggest issues right now are clockspeeds. Their 14nm chips just aren't hitting high enough frequencies to really be competitive, while Pascal is near 2GHz easily.
It's all about price though. A GTX 1080 Ti runs $700 , a GTX 1080 runs $500 or so typically.
If they can compete with the GTX 1080 and be slightly behind the GTX 1080 Ti but offer the card at say $400-450 they'd get many more sales than offering better than GTX 1080-level performance for $550-650.
The flagship must
be faster than the GTX 1080, however. It's much too late for any full die card to compete with a <$350 GTX 1070 while still making AMD decent money.
Whatever the cut down flagship is, assuming the same ratio as the R9 Fury vs R9 Fury X it would be 3584 shaders ; this would likely be the card competing with the GTX 1070 and possibly the GTX 1080. I figure without any major improvements ~ 3500 shaders @ 1200MHz should be able to compete with the GTX 1070 so I'm cautiously optimistic (The shader increase from GTX 1060 6GB to GTX 1070 is +50% ; +100% from GTX 1060 6GB to GTX 1080). I hope that without pushing the voltage to astronomical levels that it would be at least 1300MHz base clock though.Edited by AlphaC - 4/29/17 at 4:01pm