R9 390x is looking like a dual GPU now. Remember when AMD was talking about how dual GPUs will be beneficial to VR? 1 GPU per eye. I am not sure if I like this actually.
Fudzilla ran an article where they connected the dots that R9 390X must be a dual-GPU card because AMD's slides showed that VR runs better with multiple GPUs. That article made 3 flawed assumptions:
1) It assumed AMD didn't make a VR presentation with a single card
2) It assumed AMD didn't have 2x R9 390 cards in CF when making the presentation.
3) It assumed the maximum HBM1 VRAM size is 4GB per single GPU.
But let's stick to the dual-Tonga XT theory in the R9 390X:
First reason the theory doesn't align is there is way
too much memory bandwidth when CF doesn't work which results in a highly inefficient use of expensive HBM memory. You have 640GB/sec memory bandwidth via 1.25Ghz HBM, so let's say 320GB/sec per each GPU. However, Tonga has 40% higher memory bandwidth efficiency than Hawaii/Tahiti. That means its 176GB/sec is equivalent to 246GB/sec on a 7970/290X (!). Despite so much memory bandwidth, Tonga can't even outperform an R9 280X. Why would you give each such GPU 320GB/sec bandwidth+40% memory bandwidth efficiency by using a very complex and expensive HBM just for the sake of lowering power usage? You would be wasting a ton of memory bandwidth. When CF doesn't work, you'll be wasting MORE than 50% of that 640GB/sec memory bandwidth since a 2048 SP Tonga with 32 ROPs and 128 TMUs is too slow to take advantage of it.
What you'll end up with is a $650 card that loses to a 1 year old R9 295X2. Since today consumers choose a GTX980 or a Titan X over a $650 R9 295X2, this strategy would be an instant fail.
Second reason this theory doesn't make sense -
perf/watt. AMD would be better off then just taking R9 290/290X and slapping 2 of those on a single card with an AIO CLC + HBM because Hawaii has superior perf/watt than Tonga. So if AMD was severely strapped for time and cash, why use an inferior mid-range chip to start with?
That means if AMD were going to take full advantage of dual-GPU setup + HBM, why even bother with Tonga? Since they already managed to design R9 295X2 with 1Ghz 290Xs on there, might as well add HBM to Hawaii since R9 290/290X has better perf/watt, while that 120mm AIO CLC cooler can easily cope with 500W of power usage. At least with this solution when CF fails, you get
92% of 980's performance. But most importantly, AMD's new strategy is not to use mid-range performance chips in a dual-GPU solution (i.e., AMD themselves never made 5850X2, 6850/6870X2, 7870X2, etc.).
Third reason this theory doesn't make much sense -
Tonga's performance when CF wouldn't work.
You see R9 290X is 59% faster than an R9 285. Even if the full Tonga XT is 10-15% faster, when CF ultimately doesn't work, you'll end up with R9 390X losing to R9 290X by a whopping 30-40%. AMD would never release a next gen GPU for $700 that would lose to an $280-300 R9 290X by 30-40% in modern games where CF isn't supported.
Forth reason - the dual-GPU card using mid-range size chips vs. NV's flagship monolith die has failed every generation. Why would AMD repeat the same unsuccessful strategy? 5970 vs. 580? Ya, that didn't work.
Fifth reason - the future. If you say R9 390X is made of dual Tongas, then the best single chip remains Hawaii. Guess what that means? That means for 14nm generation, AMD will be asked to make a chip at least 2.25X faster than R9 290X in 1 generational jump. Titan X (x1.5) and Pascal (at least x 1.5 over Titan X) = 2.25X increase over R9 290X. It makes more sense to design a way larger chip now to have the experience necessary to make the transition to 14nm large chip easier. AMD knows it has to increase die sizes sooner or later, might as well try it NOW on a very mature 28nm node that is more forgiving than a brand new 14nm node.
Sixth reason - dual-link HBM1 means we do not need 2 chips to have 8GB of HBM on 1 card. When Fudzilla published that article on 390X being made up of 2 chips, they were oblivious to the existence of dual-link HBM1.
-----
Now think about the die size:
R9 290X has
37.5% more TMUs, SPs,
100% more ROPs than an HD7970 despite a die size increase of 24.4% (438mm2 vs. 352mm2).
550mm2 28nm would be a 25.6% increase in die size. That means AMD could squeeze 37.5% more TMUs/SPs, and again double the ROPs. However, hold that thought. Remember how R9 285 has
higher pixel fill-rate performance with only 32 ROPs against a 64 ROP R9 290(!)? That's right - Hawaii with 2X the ROPs loses.
That means, if AMD implements the pixel fill-rate/colour fill-rate efficiency improvements of Tonga into R9 390X, it doesn't need to double ROPs from 64 to 128. In fact, AMD would be fine with 64-96 ROPs. That means, at 550mm2 die size, they will have more room to increase SP/TMUs by
more than 37.5%. So we could end up with a 4096 SP / 256 TMUs / 64-96 ROP 390X at 1-1.05Ghz clocks and none of the CF scaling issues we would have with Tonga XTs.