Something on the back of my mind that resurfaced while I was looking for stuff about Turing.

One of the major architectural changes was that NVIDIA made it so Turing can do FP and INT operations at the same time, rather than one or the other. However, this had an effect of making shaders that were either FP or INT. Pascal had 128 shaders per TPC which were FP + INT. Turing has the same 128 shaders per TPC, but they're split between 64 FP and 64 INT.  In NVIDIA's whitepaper on Turing they claimed that on average in the games they've sampled there were 36 INT operations per 100 FP operations. This is about a 26%/74% split between INT and FP.

Looking at the RTX 2060 Super and the GTX 1080, two cards that have similar performance, not only does the RTX 2060 Super have fewer shaders (2176 vs 2560 on the 1080), it's also specc'd to run slower. And since half of those shaders are dedicated to INT, that means the RTX 2060 Super only really has 1088 FP shaders to work with. Even if we pretended that for the 1080, 74% of the shaders were doing FP and 26% were doing INT (I don't think this is how it's really partitioned), that's still only 1894 shaders doing FP work.

So either 1088 shaders are doing the work of 1894 shaders, or something else is up.