I think scary that iPhone15Pro uses the world's most advanced 3nm process,
but it is only 10% more powerful CPU and 20% more powerful GPU than 4nm (=5nm) of one generation ago.
It means the moore's law is dead after 3nm makes difficult to develop Switch3.
But it is too early to worry when Switch2 is not released yet...
To be fair, N3 is basically TSMC squeezing every last drop out of FinFET to buy time for them to get the transition to GAAFET right. So expectations were (or at least,
should have) always been a smaller leap compared to the moves to N7 or N5.
Have heard it said - in here even - that emulation won't even be possible going forward because of the DLSS and RT on top of Nintendo's new security measures that have made it impossible to crack the V2 which is why all games need to be jacked through the V1, so NG Switch is probably locked down to the point it'll only be able to run on itself
Nah, it won't be impossible. It's not a problem whose computational complexity is infinite; it's still finite.
...whether current or near future consumer PCs are up to it is a different consideration
Hmm, taking a look at the minimum and recommended specs for Yuzu are...
Minimum CPU: i5-4430 or Ryzen 3 1200. On Intel side, that's 4c/4t Haswell, which introduced AVX2. On AMD side, that's 4c/4t Zen, which does support AVX2, but FP/vector throughput is 2x128 bit, so that reduces AVX2 to half speed relative to Haswell-and-later (and Zen 2-and-later)
Recommended CPU: i5-11400/Ryzen 5 3600. Mainly bumps things up to 6c/12t, as they do recommend at least 6 threads.
Optimal's more or less 'get the fastest gaming CPU'
Unfortunately, it's hard to tell how much headroom there is for single thread power here, which makes it even tougher to guess how much more would you need per thread for the NG. Although the recommendation of 6 threads or more is interesting for a 4c/4t device. That does leave me wondering if a 6c/12t CPU would be sufficient for emulating the NG
RAM: minimum (with discrete GPU) of 8 GB, minimum (using integrated graphics) of 12, recommended of 16, optimal of 32.
If RAM needs scale up proportionally (
not necessarily the case), we're looking at a what, at least triple? So minimum of 24 GB with dGPU, 36 with iGPU, recommendation of 48, and optimal being 96?
Currently existing builds probably wouldn't meet the recommended bar, no. But that bar won't be hard to clear in the near future. A pair of 32 GB DIMMs/sticks in a new build, easy.
Aside: "But a pair of 32 GB is still kind of expensive?" Yea, for now. But that's why you pay attention to
relevant news. I'll put the explanation in a footnote.
Discrete GPU: this is where things get a bit interesting...
Minimum's basically asking for 4 GB VRAM, recommended's asking for 6 or 8, and optimal's asking for 10 or 12.
... but there do exist outliers.
Certain outliers with strong demand.
Read
this.
So, I'm alluding to something. If you've been reading what I've posted for a while, you can figure out what. But here are the steps:
See the Micron roadmap I linked to above. Think about what the VRAM situation for discrete GPUs might look like in the near future.
Think of what the biggest games on the NG might end up being like. Like the next open world Zelda. Think of what might need to be done, and how that might not translate so cleanly to PC GPUs, as that Yuzu blog post discusses.
It's up in the air as to whether a (early?) GDDR7 era high end consumer dGPU will have sufficient memory to avoid noticeable quality loss on
certain outliers.
*so, DIMMs/sticks of RAM. If I understand things correctly, a consumer DIMM contains either 1 or 2 ranks. A rank is a set of 8 memory dies. So the current consumer 32 GB sticks contain 2 ranks using 16 Gigabit/2 Gigabyte dies. Once 32 Gigabit/4 Gigabytes start getting made, you can make a 32 GB DIMM from 1 rank; that is, 8 dies. So, cost should go down. Alternatively, this should enable 64 GB DIMMs (2 ranks, or 16 dies, x 4 GB per die).
Oh, I
know some of you reading this notice that I didn't mention everything about ranks. Shush you
If you know, you know, and if you don't, that's fine, cause it's min-maxing that's not relevant to the general audience.
