How Will be Switch 2 Performance Wise?

[HoloDust]

Zippy6 said:

HoloDust said:

I'm not sure how much DLAs are important for upscaling (I don't think they are), tensor cores are what's important for DLSS - 2 NVDLA units where in T234 cause they had purpose there.

SW2 has 48 tensor cores - 2050 mobile indeed has 64, it's just Richard (DF) running everything at 750Mhz to try to emulate SW2 GPU (it should've been 755MHz, but it's close enough).

Anyway, looking at that 720p Native to 1440p DLSS Performance (which is 720 native), and then doing some math comparing how 1080p vs 1440p DLSS Performance scales from official numbers on Ampere...yeah, again comes out as 14-15% for 1080 DLSS Performance.

Well DF seem to think the DLA would have helped. They said this in 2023 before they confirmed recently that the T239 doesn't have it.

"DLSS isn't a 'free lunch' and the Tensor cores in the GPU alone can only do so much. However, if T239 includes T234's Deep Learning Accelerator, that could drastically reduce DLSS's overhead." https://www.eurogamer.net/digitalfoundry-2023-inside-nvidias-latest-hardware-for-nintendo-what-is-the-t239-processor

I didn't realise the T239 had cut back on tensor cores as well, I thought it had the same 64 as the T234 but looks like it does only have 48. So yeah DLSS on this is not going to be fast for higher resolutions at all. This might explain why Nintendo is opting for 1440p native on some of their titles and not using any DLSS.

Honestly, I highly doubt it - NVDLA is for completely different purposes and, and if we're completely honest, Richard is not that tech savvy (as I said somewhere in the thread, they should really consider hiring someone who knows the stuff, preferably with coding skills, to up their game to next level).

Anyway, 2050/3050 don't have NVDLA, nor any of RTX Ampere (or otherwise) consumer based cards - that is reserved for their professional line.