Doing a direct comparison between the mobile 2050 (which is Ampere) against the Switch 2's T239 (which is modified Ampere) while only accounting for the differences in core count and theoretical clocks ignores, among other things, the difference in GPU architecture (there was some backporting of features from newer GPU generations), the difference in memory bus width and clockspeed, the differences in power/voltage/temperature curves, the difference in API and OS overhead, the difference in DLSS model (the Switch 2 uses a lighter-weight model), and for that matter even the difference in core config (it's not just "1536 versus 2048", the scaling factor of the CUDA cores is not necessarily the same as the scaling in shader processors, texture mapping units, render output units, RT cores, and tensor cores). There's so much they didn't account for that their comparison is meaningless.
Take the RTX 2050 mobile and RTX 3050 mobile, for example. They are actually both the GA107 die, and both have 2048 CUDA cores. So with the same CUDA count, you should be able to compare them directly only accounting for the clockspeed difference like Geekerwan did, right? But the RTX 3050 mobile has half as many RT cores and a quarter as many tensor cores, but it has 2x the memory bus width, 86% the memory clock, and 1.8x the max TDP. And who knows how the voltage/clockspeed curve or thermal throttling is different. So you can't compare them like this at all!
That's my point. Geekerwan's benchmark relies on the assumption that the *only* thing that affects performance is the CUDA core count and GPU core clockspeed. And that's an invalid assumption.