OK, here are my educated guesses about the upcoming lineup.
Prediction: All new chips (except the OEM leftovers and bottom-of-the-barrel crap) will be on Global Foundries 28nm SHP, instead of TSMC.
Rationale: AMD's current TSMC lineup isn't competitive, especially in terms of performance per watt - the one area where Maxwell has absolutely crushed them. The switch from TSMC to GloFo made a big difference with Beema/Mullins; AMD claims
38% lower leakage on that chip's iGPU. Even if the difference isn't that big on desktop-grade GPUs, there's still likely to be a reasonable gain. And AMD is required to buy GloFo wafers under the "take-or-pay" agreement. Kaveri/Godavari is hardly selling like hotcakes, all of the FX chips are both years outdated and on the older 32nm process, and Carrizo is a small chip that may have some niche success, but can't account for $1 billion worth of wafers annually. This leaves discrete GPUs as the only place for AMD to fill their GloFo orders. (And Fiji is way too small a volume to do it alone.) They can pay some money up front to port the designs, get something reasonably competitive until late 2016, and meet their contract with GloFo. The alternative would be to stay with unmodified TSMC rebrands, sell uncompetitive crap until late 2016 (continuing to lose mindshare and market share), and then pay penalties to GloFo under the WSA. I don't think AMD will be stupid enough to try that, though they have done some stupid things in the past.
Prediction: All new chips will have an updated UVD block, similar to that in Carrizo, which will support HEVC decoding. They will also support HDMI 2.0.
Rationale: It is known and demonstrated that
Carrizo will have HEVC decoding built in. AMD has traditionally done well in the HTPC market, and a HEVC decoder on discrete cards would obviously be desirable. It would give AMD a leg up over Nvidia, which
only has this feature in the GTX 960 (even the Titan X is lacking it). A few months ago, on the DevGuru boards (which have since been rebooted), an AMD employee confirmed all 2015 GPUs would have HEVC support. If AMD has to go to the effort of porting the GPUs to GloFo and incurring all the fixed expenses (tapeout, validation, etc.) then updating the UVD block probably won't be much extra expense on top of that, especially if these chips are designed in a modular fashion using automated or semi-automated tools (which I think they are).
Predicted chip lineup:
- Fiji: Pretty much what everyone has already deduced from the leaks. 4096 GCN shaders with 4GB of HBM on an interposer. AMD's public statements seem to indicate that it will be 4GB, though it's possible that this could be misdirection and that an 8GB model could be released either immediately or at some point in the future. My money is on 4GB at this point, though.
- Grenada: Hawaii respun on GloFo. There's a good chance that the Double Precision support will be cut down to reduce the transistor count. Although I would bet against it, it's at least conceivable that they may increase the shader count to 3072, and/or update the architecture to GCN 1.2.
- Antigua: Tonga respun on GloFo. I don't see many functional changes being performed here except the addition of HEVC decoding.
- Trinidad: A replacement chip for Pitcairn, probably with the same 1280 shaders, but redone from the ground up as a full GCN 1.2 chip. Another possibility is that this might be a "half-Antigua" the same way that Nvidia's GM206 is almost exactly half of a GM204. The only issue is that doing so would put Trinidad behind full Pitcairn in some benchmarks (1024 shaders vs 1280). Still, that could save some money and would still be a competitive product at $99-$149 if the perf/watt was good enough.
Predicted card lineup:
- Fury XT: The top-tier water-cooled card that has been teased recently. At $849, this will go up against Titan X with approximately 5% better performance. Nvidia will focus on their brand equity and higher VRAM count in competing against it.
- Fury Pro: Cut down to 3584 shaders, but still sporting 4GB of HBM. It will have to be priced at $499 to compete against the GTX 980 Ti.
- R9 390X: Full Grenada. TDP (maximum power limit) will be 250W, although it won't go much above 200W during normal gaming. Due to higher clocks and/or architectural updates, performance will be slightly better than GTX 980 (maybe 10% or so), and the 8GB of RAM will bloat the bill of materials, so it will sell at $399. I suspect there will be no reference card, because AMD doesn't have a decent blower and they don't want the bad reviews they got with the Hawaii cards. There's an outside chance that those old shroud pictures which we though were for Fiji were actually meant for a water-cooled edition of the R9 390X with an Asetek CLC.
- R9 390: Cut-down Grenada with 2560 shaders (or maybe 2688 if Grenada has a total of 3072). Same TDP. Designed to compete against the GTX 970, so it can be priced no higher than $299. It may well only have 4GB of VRAM, since for most users this will be enough.
- R9 380X: Full Antigua with 4GB of RAM. TDP (maximum power limit) will be around 175W, and it won't go over 150W during normal gaming. Basically, better than Tahiti performance at Pitcairn TDP levels, with all the GCN 1.2 features. At $249, this should easily beat the GTX 960.
- R9 380: Cut-down Antigua, basically the same as the R9 285 but with much better perf/watt, both because of the new process and because the current 285 is trash silicon. One power connector, so the maximum TDP is 150W. Won't go over about 125W in normal gaming. May be available with either 2GB or 4GB of VRAM and should have a price around $199.
- R9 370X: Full Trinidad. 100W TDP, designed to beat the GTX 750 Ti in performance (though not perf/watt) at a similar price level. With 2GB of VRAM this should cost no more than $149.
- R9 370: Cut-down Trinidad (or maybe just downclocked?) with a 75W TDP so that it requires no external power connector. This will compete in the same market that the GTX 750/750 Ti is currently in. If AMD can hit a shelf price of $99, then it would be a tempting upgrade for people who want to add gaming capabilities to an OEM system. HEVC decoding would be especially nice to have in these kind of systems, because they may not be strong enough to do it on the CPU alone.
What do you think? Does that sound reasonable?