Sir, where did you find this info? Cache Block diagram looks exactly like the one from sandy bridge times, everything seems to be the same as inside kaby.
Nope, the structure is the same, but the amount increased. This leads to better mining performance in cache-bound algos, and better flexibility (w/o looking at core count). E.g. on i5 8600K you may now use 2 double threads for cryptonight (2x4Mb) at around 320+ H/s and you will be able to use the rest (1Mb) for video decoding (which uses cache as well). This is good, and affects power consumption and heating (in a good way) as well.
It's the same 2Mb per core on i7 and 1.5Mb per core on i5, just more cores. So no real changes
There's no chnage in architecture that I'm aware of, everything seems to be scaled 1.5 vs previous generation,
no compile differences, no source code differences.
It's not a SW issue, I'm doubtful it's a Coffeelake issue. It might be an issue with that specific CPU or system.
More data is required including from other Coffeelake owners.
OP has been quiet for a couple of days, maybe it's no longer an issue.