I think $300 is feasible. If we're Q419 or Q120, I think we're 10/12nm instead of 7nm. I think if we're taking
Thraktor 's statement into account, then we should be looking at A73 instead of A75 or A76. I don't know how much LPDDR4 or 4x costs. The best guess is $40-$80 depending on type and speed. I really wish I had a better number for this. As far as memory bandwidth, I think if we're talking 12nm, we're probably on Xavier (probably easier to modify a 12nm design than modify and shrink a 20nm design) and according to
wikipedia Xavier has 137 GB/s of memory bandwidth.
Sorry this is so discombobulated.
If were talking Q420, then we're probably safe talking about 7nm.
The CPU choice is actually kind of interesting. ARM splits their cores into high-performance and high-efficiency designs, but for a portable console you actually want something in the middle, that can provide as much performance in as possible in a fixed power budget. Going from A57 to A72 to A73 there's definite improvements for a use case like Nintendo's. The A72 is an updated version of the A57 design (itself based on the 32-bit A15 core) which improves in pretty much every aspect, as it's more powerful, more power-efficient, and takes up less die space (ie it's cheaper).
The A73 is actually an evolution of the 32-bit A17 core, and it's quite a different approach to the A57/A72, with a shorter pipeline and narrower front-end. This leads to a smaller, more power-efficient core. It's actually about 35% smaller than the A57 on the same process, and a 12nm implementation could be less than half the size of the A57s on the 20nm TX1.
The A75 is a revision of the A73 based on improving peak performance, and ARM didn't actually claim any efficiency improvement over the A73 (which holds up in real-world testing, as the Snapdragon 835 and 845, running A73 and A75 core respectively both on 10nm processes put in pretty similar power efficiency figures). For a new Switch, this increased peak performance wouldn't make much difference, as Nintendo wouldn't have the power budget to hit A73's limits anyway. It's possible that the A75 is close enough in die area to the A73 that it's worth using the newer version anyway, but I wouldn't expect any meaningful difference in performance between the two in a fixed power budget.
The A76 is a more interesting one, as it's a completely new core design, and in actual products, it's a big improvement in both performance and power efficiency over older A75-based designs. The problem is that these new SoCs are built on a 7nm process against 10nm for their A75 predecessors, so it's very difficult to tell if the efficiency improvements are all from the process change, or if there's some gains from the architecture, too. ARM themselves don't make any ISO-process efficiency claims (like they did for the A72 and A73), so I would assume they took the efficiency win from 7nm and focused on performance.
So the reality is that using the A73 or A75 or A76 probably won't make much difference to Nintendo if they're limited to ~500mW per core, so they may as well go with the smallest (ie cheapest) core, which is the A73.
There is one other factor, though, which is Nvidia. For the first Switch model Nintendo used an off-the-shelf SoC designed by Nvidia, and although there isn't an off-the-shelf option for them now, and I'd wager good money on Mariko being designed primarily for Nintendo, there's a possibility it's not designed
entirely for Nintendo. They may have entered a deal where Nvidia can sell on the Mariko chip to other customers (let's say they want to make a play at the Windows on ARM market), and in that case Nvidia would want an SoC with those higher peak performance levels you get from the A75 or A76.
