JEMC said:
It's not as easy or simple as you think. AMD has only stated that they'll start shipping thir Polaris GPUs between June and September for the important "Back to School" season, and portable GPU are expected to come first. We still don't know when desktop parts will come. The 14nm manufacturing process (I'll leave the competing 16nm from TSMC aside as AMD won't use it for Polaris) is still fairly new to the point that AMD has split supplier between Global Foundries and Samsung. That's because the production volume is still not high enough to guarantee a good supply with only one source and also because while the process efficiency is acceptable enough to start mass producing them, that doesn't mean that it's as good as they would want it to be. You are also forgetting that while NX could launch in November (that's when Nintendo usually launches its consoles), they will have to start producing them weeks before to have at least a certain amount to meet the demand at launch. Let's not forget the Digitimes report where they stated that Nintendo was targeting shipping 20 million NX in 2016. Even if they only reach half that production, that means months of production, not weeks. Lastly, while you are completely correct saying that costs are based on how many chips you can get off of one wafer and that with 14nm you'll get more chips per waffer, you are forgetting two important things. 1- right now the process is not efficient enough and that the percentage of faulty chips is a lot higher than the 28nm counterpart; and 2-You are assuming that the 14 and 28nm cost of producing a waffer is the same, but it's not. 14nm demands better equipment and is considerably more expensive than 28nm, so even with the extra chips that they can get from it, the cost per chip is higher than a similar size chip at 28nm. Oh, and that 40 CU chip from AMD is just speculation. We know from a leak that there will be a Polaris 10 chip with 36 CUs, and some are guessing that there will also be a higher version with 40, but nothing has been confirmed (or denied). And we don't know if that chip will replace the 290/390 series or if it will improve their performance. |
Actually it is that simple, for one thing Polaris 10 is confirmed as the line for the high end desktop market, 11 is the mobile and smaller desktop line, but all use the same Polaris architecture.
Just because Nintendo are targeting 20 million units shipped by the end of Fiscal Year 2016 that doesn't mean they need all of those units at launch or even for this year. A few million units is plenty for launch, plus you have to remember that a wafer could produce way more SOCs at that level of performance for a Handheld than if they were used for high-end GPU chips.
Supply targets are only targets, tbh it's not reasonable that Nintendo would think they could sell 20 million units by even June 2017, not if they plan to release this Holiday. 20 million could maybe be sold before the Fall of 2017, if the NX handheld or whatever it is, was attractively priced to the point where they're flying off of shelves like Wii or PS2 did.
When Nintendo says they're targeting to ship by the end of 2016, they mean Fiscal year end, which is March 2016. If Polaris is available between July and September of this year, then Nintendo could start production when that begins. If Foxconn makes and ships 1 million NX devices a month from August through to November that's 3 million units available to sell by November.
Foundaries can provide higher chip volumes and Nintendo can ramp up volumes of consoles as we enter 2017 (calendar year), to increase shipped units to their target by the end of the fiscal year. Monthly quantities don't have to be the same, they can increase in output each month.
It's possible that AMD split between two foundaries, one for their own chips that they will sell to the consumer market for GPUs and the other for semi-custom parts (admitedly speculation, but logical, since these are different kinds of chips, with one incorporating CPU and GPU on one die).
My point about the use of Polaris for a handheld is that it fits the bill, so would a 14nm CPU with better architecture than AMD's Puma, it works with the power consumption necessary to fit in a handheld device like the one mentioned in that rumor, though it should of course still be taken with a grain of salt because it's a rumor.
As far as costs go, I was going by AMD's own target costs for the part they said they're aiming to be on par with their own R9 290x, they said they want this come under $349, releasing this back to school period. A 1.3TFlop part would need way less space than an 4TFlop one. I said 40 CU because AMD said themselves that they wanted to release a Card with the same performance as the 290X, which has 40CUs, unless with the newer core tech they've also improved performance per CU, without having to increase clock speed (which is possible, since there are a lot of changes to each GPU Core).
Wafer costs don't have to be the same if the volume of dies produced is greater than the 28nm die, with defects taken into consideration. AMD can still make more per 14nm wafer than on 28nm and produce cost competitive chips, but that would depend on the real production figures, which I don't think are public.
