In The Callisto Protocol, my ZhiTai TiPro9000 shows high sequential speeds, but the 1% Lows are all over the place. Is this due to write amplification?

Test Report CP-ZZ in a Windows 11 23H2 environment with a custom kernel used 3DMark Storage Stress and found read sync rates oscillating between 4.1GB/s and 5.2GB/s, with a momentary peak of 7.1GB/s. I mistakenly chased a driver bug for days, only to realize the culprit was a forced cache reclamation logic triggered when the SLC buffer dried up. By adjusting the prefetching size in low-latency mode, the frametime curve flattened out significantly, and the perceivable stuttering was slashed. That said, during massive fast-asset streaming, I still faced minor jitters in the 2ms to 5ms range. Actual data showed a deviation of only 3% from manufacturer specs, confirming the hardware is fine, but driver-level optimization is stubbornly limited.