I mean dlss runtime is dlss runtime. As far as I understand there's nothing console specific optimizations can do to mitigate that, unless they're degrading quality.
Sure, but let's take Rich's own words into account here
This is the closest approximation we can get for the t239 GPU but, more accurately, what you're going to be seeing is an ultra low-spec ampere GPU running at low clock speeds starved of memory bandwidth. This is only going to give us a ballpark idea of what a mobile Ampere chip can deliver
How big a ballpark are we talking here? Just spitballing, but he's using different hardware that he's manually underclocked. If he gets with 10% of the real power of the hardware, I'd call that a slam-dunk.
He's measuring DLSS cost inside of a video game, not in a pure test case. He's comparing the
average frame time of a native res run to the
average frame time of a DLSS run. That's going to have at least a tiny bit of error. Plus
Death Stranding's DLSS implementation is a black box, where other settings may change along with enabling DLSS (like post-processing). Getting within 5% would be impressive.
Rich's measurement is 18ms. If we apply that combined 15% fudge factor, we get a range of 15.3-20.8ms.
The biggest conclusion is that DLSS 4k is expensive, not "possible" or "impossible." Both of those numbers are certainly within the "viable for 30fps" range, at least.
Whether or not that level of DLSS is worth it is another question, but leave that to game developers. Which brings me to a second Rich quote.
Remember that Switch games typically receive some tweaks to content in order to smooth out performance, or we get stuff like Dynamic resolution scaling added. It's called optimizing for the strengths and the weaknesses of a specific platform, and that's a magic ingredient that none of our testing here is going to be able to highlight
We can't lower DLSS time, but can we find ways to
hide it on console? And the answer is yes, Nintendo already has an optimization that it uses on Switch that would be very handy - NVN lets you render one frame while the CPU is busy with a different frame. Instead of having 16.6ms for CPU and GPU
total you get 33.3 ms, while staying 60fps. There are some caveats to this, but it's a very common trick for Nintendo first party games.
Looking back at the error bars, 15.3ms seems positively reasonable now! Which brings me to a point that is totally personal. I think Rich has absolutely nailed it, but
I don't believe Nintendo would choose to make DLSS 4k just barely impossible.
Nintendo choses the hardware they get. Rich's tests are very good, but T239 wasn't a chip found in a dessert. If DLSS were half as fast as it seems to be, sure, Nintendo probably wouldn't go out of their way to make it work. But if it's close, Nintendo would be crazy to not go out of their way to make it possible.
Perhaps that means a slightly bigger fan and pushed clock speeds. Maybe it means a custom "GPU overdrive" mode, where CPU gets clocked down to nothing, and the GPU gets overclocked. Rich's overall point is that 4k isn't free, and we shouldn't expect it universally, which I think this thread already understood when that video came out. But if it's nothing more than a little bit of tweaking to get there, I think Nintendo would make sure that they had all those tools in their pocket.