Would the graphical jump from XBO level to PS4 level, really make a difference in the downporting to handheld format? This seems rather trivial. I mean, if you want do downport the latest Assassins Creed (just for instance)... would it really matter if it came from XBO or PS4 level graphics?
As for the other two arguments, i guess the primary question is, how expensive (or cheap) is the tech now, 3 years later. So i was thinking about that, if you look at it from the same perspective as the CPU.
It's more a matter of Nintendo's internal games. Assassin's Creed is going to be difficult (or impossible) to port to a handheld regardless of how powerful Nintendo's home console is, as it's going to be built around the highest possible target (and in Unity's case a higher target than any console could reach). For the next Mario, Zelda, Splatoon, or whatever else, if Nintendo wants to develop it in tandem for a 1080p console and a 540p handheld, then the easiest case is a console which shares the same GPU architecture as the handheld and is roughly four times as powerful. They would have to use lower-res textures (trivial) and simpler models (easy if you've got the tools), but they could use the same graphics pipeline, the same lighting techniques, the same shader code, etc. Effectively if you made a game which runs at 1080p and 30fps on the home console, you could be pretty confident that it would easily scale down to run at 540p and 30fps on the handheld (unless it's CPU-limited, which is a more difficult issue to deal with).
The further you move from that ratio the more difficult (and more expensive) it becomes to develop a game for both. There isn't a sudden step change where it's easy at 1.2 Gflops but expensive at 1.3 Gflops, but it gradually adds to Nintendo's development costs, and deteriorates their attempt to have two devices share the same library, the further they push up the performance of their home console. If they are fully pursuing the shared library concept, then they may decide that roughly XBO-level performance is a happy medium for them; it would bring them just within the scope necessary for third party games while keeping their own cross-device development feasible.
A Polaris 11 variant would be a good fit in this range if they pushed hard for 14nm. Bristol Ridge is rumored to have a 16 CU GPU part on its APU at 28nm as well, though it seems to be speculation. It would only make sense as a semi-custom part since it would still be limited by DDR4.
I'm with Blu in believing that 14nm and 16nm are extremely unlikely for a 2016 home console. By most accounts we won't be seeing a full changeover of PC GPUs to the new processes until 2017, which points to a slowly maturing node (by comparison both Nvidia and AMD had fully switched over to 28nm within a few months of the node's existence).
Basically, I can only see two reasons Nintendo would go with 14nm for the NX home console
1. They want to significantly outperform PS4, and effectively get an early jump into next gen performance rather than a late jump into this one.
2. They want to use exactly the same GPU architecture for both home console and handheld, and pre-finfet AMD GPUs would have been incapable of achieving sufficient perf/W for the handheld, so they have to use 14nm Polaris for both.
The first option is not only very unlikely, but simply wouldn't be possible for an affordable price this year. In 2017 we could potentially see something with double PS4 performance on 14nm for a somewhat reasonable price point, but I can't see it happening this year.
The second option would be theoretically possible this year, if we were looking at perhaps XBO-level performance, but would likely be more expensive than the same performance out of a 28nm chip. In fact if they did take this route I would be very surprised if it was any more powerful than XBO for the reasons above; it wouldn't make much sense to spend extra on your home console SoC for parity's sake only to have a massive differential in performance between the two.
Very interesting, especially the part about the 4 chips vs 1 not being the major factor in why HBM is expensive. Do you think 4GB HBM+1 6GB LPDDR4 chip on a 128bit bus (2GB reserved to the OS and 8 total for games) is feasible for a 299$ console? Or at least 2GB HBM+6GB LPDDR4?
It's hard to say. To be honest for $299 I think the best we'd be looking at is 8GB of either GDDR5 or LPDDR4, but I suppose that depends on the expense of other components and Nintendo's willingness (if any) to take a loss. The packaging costs of using HBM will certainly come down over the next year or two, but I'm not sure how much HBM1 will come down in cost with HBM2 coming along so soon to replace it.
In fact, I think that it would be worthwhile to consider likely cost reductions over the life of the console in addition to just the launch costs. The Wii U had almost no scope for cost reductions, with a GPU made on an already mature 40nm node without scope to shrink (due to the eDRAM), an MCM package that is unlikely have come down in cost much, a small CPU which could have been shrunk but for trivial savings, DDR3 which was already mature and commoditised, etc. This may be good reason for Nintendo to favour technologies which have clear scope for future cost reductions this time around.
HBM1 would certainly see cost reductions on the packaging side, but perhaps not a whole lot on the chips themselves (having only one vendor wouldn't help Nintendo on that front). On the other hand, if they used two 1GB HBM1 chips on a 2048-bit bus in the launch console, it may be possible for them to switch to a single 2GB HBM2 chip running at twice the speed on a 1024-bit bus when they shrink to 14/16nm. DDR3 and GDDR5 are unlikely to see much in the way of cost reduction, whereas DDR4, LPDDR4, GDDR5X and HBM2 should all come down in cost quite a lot over the next few years. Using industry-standard tech which is currently in the growth phase means you have ample opportunity for cost reduction over time, even if it means spending a little more at launch.
On a similar front, I put together a quick little table of potential RAM options for a late 2016 console (obviously some of them more likely than others). The specs are all
per chip, and the higher speeds aren't always available on the highest capacity. They're ordered roughly in increasing order of power consumption:
Code:
Max capacity Max speed Max I/O Max bandwidth
LPDDR4 6 GB 3733 MT/s 64 bit 29.9 GB/s
HBM2 4 GB 2000 MT/s 1024 bit 256.0 GB/s
HBM1 1 GB 1000 MT/s 1024 bit 128.0 GB/s
DDR4 2 GB 2666 MT/s 16 bit 5.3 GB/s
DDR3 2 GB 2133 MT/s 16 bit 4.3 GB/s
GDDR5X 1 GB 12000 MT/s 32 bit 48.0 GB/s
GDDR5 1 GB 8000 MT/s 32 bit 32.0 GB/s
This is just based on the publicly available information I could find, so any corrections would be welcome.