I'm amazed that Apple still can't cut out arch-nemesis Samsung from the picture. Good game, TSMC.
Apple and Samsung have fought bitterly in the courtroom when it came to patents, but the two firms are still prepared to do business: namely the buying and selling of chips. Earlier this year, Apple reportedly decided to drop Samsung as its manufacturer of A-Series system-on-chip processors used in iPads and iPhones. Instead, …
TSMC is extremely capacity limited, despite being by far the largest foundry on Earth, because almost everyone on Earth goes to them to have chips fabbed.
Apple needs to have a backup plan in place in case they can't get the chip volumes they require from TSMC. I suspect they'd prefer to use TSMC for the A9, and surely will do a design that targets both TSMC's and Samsung/GF's processes, so they can switch sources as needed if one runs into any production delays or capacity constraints.
Going to TSMC probably caused a lot of pain for Samsung as they were left with a lot of excess capacity. Apple may have got a very nice discount to win their business back.
You say it yourself:
"TSMC is extremely capacity limited, despite being by far the largest foundry on Earth, because almost everyone on Earth goes to them to have chips fabbed."
The practical reality was that Apple was facing not getting anywhere near enough of the chips fabbed that they wanted within the time-frame they wanted them. Samsung could afford to smile knowing perfectly well when Apple (allegedly - we don't actually know) are supposed to have dropped the Koreans that Cupertino would in all likelihood have to think again. I would be very surprised if Samsung did not have a pretty clear picture of current fabbing capacity planet-wide, such industrial intelligence is essential for any company when deciding pricing strategy in general. They would have known that Apple simply could not afford to drop them.
The alternative side of the negotiation word have been Apple knowing that the Samsung mobile sales tanking is hurting the entire group and that Apple could fill up both some of the slack demand as a result of Sammy mobile needing to scale back production, and are more able to withstand short term supply constraints by limiting demand in various markets.
Whichever way it went I would love to have been a fly on the wall....
Where's your proof that TSMC production was limiting Apple? They seem to be selling tens of millions of iPhone 6s - they'll have their biggest launch quarter ever. As always the limitation is Foxconn, because even they can't produce at the quantities Apple needs during launch. Apple is always in shortage after the iPhone launches.
A smaller feature size doesn't necessary mean Samsung's process is better or lower power. There are a lot of dimensions to measure transistors aside from gate length, and it is well down the list when it comes to power draw. The efficiency of the FinFETs will dominate for mobile SoCs.
Apple seems to have a fair bit of confidence in Samsung's ability to work out the kinks Samsung has reportedly run into:
There are a lot of dimensions to measure transistors aside from gate length, and it is well down the list when it comes to power draw.
Errr - not really.
P ∝ f2C
C is pretty much determined by the gate dimensions - there's little you can do to the dielectric that leaves the gate tenable. So your power draw is determined by voltage, clocking frequency, and gate dimension. For a given performance, there's a limit to how far you can reduce the voltage, so power draw for a given task is pretty much determined by gate size.
 This is naturally the averaged clocking frequency, integrated wrt time across the device and divided by gate count; a static module that isn't changing state consumes only leakage power.
Leakage tends to dominate at smaller nodes, especially in lower power devices. Hence the move to FinFET. The difference between a FinFET and non FinFET device of the exact same gate length is fairly large in the 14nm/16nm range. Making the fin taller/narrower helps even further, while not changing the measured gate length.
Leakage tends to dominate at smaller nodes
Only when the device is idle. If it's switching - you're charging/discharging capacitors (since that's what a FET gate is), and the Vf2C thing dominates.
If your leakage current is greater than your switching current, your process is useless.
This is particularly true here: the only reason TSMC calls their next node 16nm and not 14 as others is because 14 is unlucky (for some reason I forgot, maybe a Chinese speaking person can comment). Everybody in the industry considers 16/14 as the same node, and if there will likely be performance variations the 14/16 number itself is meaningless.
And by the way, for TSMC it's best understood as mostly 20nm with FinFETs added. There is also a 16FF+ coming quickly after with some incremental performance and density optimizations. As I understand it's the same for Samsung/GF: the node change is justified by the move to FinFETs and the performance/power impact, not by a true shrink.
Biting the hand that feeds IT © 1998–2019