That is interesting. I am not aware of any large perf regression with POM but I can check. My gut reaction is that 500 steps is an awful lot and it can make the shader very sensitive to performance. It is possible that some unrelated change in the engine, such as a minor increase in number of temporaries for unlrelated shader options could have reduced how much the gpu can parallelize.
Also, when using very high step counts, texture formats and mip handling can make a huge difference. Just changing the height to be uncompressed can cripple perf. I will have to double check later, but its possible that 4.13 is when POM was changed from using SampleGrad to SampleLevel with mip0 specified. At some point I changed it so that it did SampleLevel because in most test cases it was actually faster and TempAA was doing a better job cleaning things up than semi-accurate mips. It may be that for some textures and materials is backfiring.