Also the stuff about Fiji being a lower end GPU with HBM (4GB) and bermuda being a higher end HBM one with (8GB), the theory that raghu at least bases his speculations on, seems to be most likely false.
From what I've read, Bermuda XT is just a codename for a dual R9 395X2 card (i.e., dual Fiji XTs).
http://cdn.videocardz.com/1/2015/04/Hot-Chips-Symposium-Fiji-2.5-HBM.png
^ If AMD had a higher end GPU than fiji and fiji was going to be used for the 380X instead of the 390X they wouldn't be holding a presentation about fiji in august. They'd be doing a bermuda presentation.
That is true. Chances are Fiji is the code name for the next gen flagship succeeding R9 290X. I think what raghu alluded to is regardless of the specific codenames, his view is that R9 380/380X cards will be a complete redesign and not improved Hawaii XT chips. For example, they could use a combination of Tonga's architecture + HBM. I personally think R9 380/380X will be improved Hawaii XT chips with GCN 1.2, updated UVD, HDMI 2.0, more mature 28nm node, but no HBM. I think HBM will be limited to Fiji chips this gen. However, the problem with my theory is it doesn't at all explain then how AMD will be able to improve their mobile dGPU line-up. If they don't make any improvements in perf/watt other than for their HBM chips, then their mobile dGPU strategy for the next 18 months is a failure as they won't be able to compete with Tonga variants in laptops.
And the rest of the lower end leaks for mobile and for desktop have showed reasonable proof for rebrands. Even the leaked XFX double dissipation card had a pcb that was identical to hawaii PCBs from everything we could actually see on the card.
But even if you have identical PCB to R9 290/290X, we have no idea what the GPU is like underneath. Therefore, that alone told us little.
Personally I think that AMD's architecture improvements over tonga are limited to nonexistent and as such it didn't make any sense to design any more GPUs which would be only slight improvements over the old stuff, but expensive to design.
We'll likely see Fiji with almost identical GCN to tonga that gets its extra performance from the power savings from HBM, added die size and added space on the die caused by the change in memory controllers.
Why do they need to make major improvements in architecture beyond Tonga? Tonga's architecture more than doubles their geometry performance, increases memory bandwidth efficiency 40%, increases pixel fill-rate efficiency 70%. At that point, all they really need to do is grow their die size to increase shaders and textures and use HBM1 to lower power usage. Why is that not a good solution? In Tonga's case, the GPU is basically shader and texture limited and because it only has 32 ROPs, it had no chance of ever beating a 290. On paper it looks like Tonga's 384-bit controller is 20% larger in size than R9 290X's 512-bit because it's the same controller used in Tahiti - highly inefficient old gen one. Most games are mostly shader and texture limited, not tessellation limited. That's why Tonga is such a poorly executed product but I believe it was just a card made specifically to test out the underlying architectural changes R9 390 will have. I still think AMD will make other improvements that we haven't seen in Tonga.
This is a stopgap generation anyway. The real deal will be when 16nm rolls out, both manufacturers are on HBM and we'll see which architecture actually is the superior one.
No doubt that next gen should be a bigger improvement in performance, at least on the NV side. 780Ti doubled 580's performance but Titan X is nowhere close to that. Pascal will have HBM2 + 14nm + new architecture so it's not out of the question the performance will be 80-100% over Titan X.
Exactly. It would be as silly as AMD going BACKWARDS on IPC from Phenom II to Bulldozer.
Oh wait.
I don't think this analogy applies in this case. The reduction in IPC with Bulldozer over Phenom II was because BD was a brand new architecture, completely different from Phenom II. There is no indication at all that R9 300 series is some new radical architecture. If anything, it will build on the existing foundations of GCN 1.0-1.2 which means it's hard to imagine how its IPC will decrease. I think more efficient ROP performance and double the geometry performance should allow those 4000+ shaders and rumoured 256 TMUs to put their power to use.
If we look at R9 290X, it should be 15-16% faster than R9 290 on average (Shaders: 2816 x 1000 / (2560 x 947mhz or TMUs: 176 x 1000 / (160 x 947mhz)), but it's nowhere close. That means there is some major bottleneck(s) in Hawaii where the extra TMU/shaders are wasted. My educated guess is the geometry performance and pixel fill-rate efficiency are holding back R9 290X's extra specs.
I keep repeating this point because there is no way around it:
918mhz 32 ROP GTX285 has 20% higher real world pixel shading power than a 947mhz 64 ROP R9 290. That tells us right there that Hawaii XT's 64 ROPs are highly inefficient. AMD needs to resolve this issue with R9 390 series because right now those 64 ROPs on Hawaii XT are worse than 32 Tonga ROPs. That's pretty shocking. Knowing this though, it shows just how much performance R9 390X can gain at 4K considering today R9 290X does pretty good at 4K despite its lackluster 64 ROPs. Imagine putting 64 "Tonga-style" efficient ROPs into R9 390X + Tonga's 40% memory bandwidth efficiency!