Nintendo Switch model is expensive for us and cheap for Nintendo. Hardware and Software.

[Pemalite]

CrazyGPU said:

But PS4 is 3 years old, besides that, it had a custom soc with the power of a mid range PC graphic card, not a tablet soc like Switch. PS4 pro just lauched this year with new AMD Polaris architecture and more than twice the graphic power built in 16 nm. And Scorpio will probably bring AMD VEGA.

And the main point is, Switch is mostly a tablet, so it really needs new transistor technology to bring power down on the go. So putting in it old Maxwell at 20 nm is not going to help. They went for cheap stuff, maximizing profits. Fanboys will buy anyway.

The Playstation 4's age is irrellevent. It is derived from half-decade old technology today.

Polaris is based on Graphics Core Next 4.0, which is an evolutionary update over prior designs, dating back... You guessed it. Half a decade.
Heck a large portion of AMD's 400 series lineup is rebadged older Graphics Core Next 1.0 and 2.0 parts destined for the OEM markets.

If Vega drops it will either be added on top of the 400 series stack or usher in the 500 series, again, mostly being rebadged parts stemming all the way back to Southern Islands/Graphics Core Next 1.0.

Tegra X1 is fabricated at 20nm Planar. Which is actually the same as TSMC 16nm, but has Finfet, Volta will be built on 16nm Finfet and double the performance.

However with architectural choices, you can reduce power consumption and increase efficiency. Two chips at the same fabrication level with the same amount of Transistors can perform vastly different and use vastly different amounts of power.

nVidia for instance takes a tiled based approach which has given them an efficiency edge over AMD since Maxwell, Tiled Based Apparoaches tend to be more effective at using minimal resources that are available like bandwidth and caches, so there is less wastage... Combine that with things like Colour Compression and the ability to execute two FP16 operations bundled together as an FP32 op and other architectural nuances and it's easy to see how nVidia manages to get such efficiency even on older fabrication process's.

And that isn't all either. Fabrication is only part of the story. - A chip built on an older and more mature fabrication node can not only be cheaper, but use less power on a newer node.

And even nodes like Global Foundries 14nm Finfet has wildy differing levels of performance/leakage, which greatly influences power consumption and clock rates.