Can't speak to your hardware, but I once had a similar situation where performance would be good after I started windows and then drop over time.
Turned out, all of the startup stuff that Windows delays and does in the background after you already have your GUI up and running was keeping my cores from idling down to a slower clock speed. The CPU involvement in drive activity was sufficient that this was enough to make a difference between early on, when most of the cores were active and thus at higher clock speeds, and later on, when they were mostly inactive and running at a lower speed.
TL;DR: If the CPU load is not over a certain level, the CPU might be underclocked, which could effectively increase the OS overhead in drive I/O.
You could confirm or refute this theory either by turning off SpeedStep or by just arranging to have a sufficient non-I/O-activity load distributed across your cores while you test throughput, whichever is easier.
Last edited by AikenDrum; 10-18-2018 at 01:14 PM.