Normally benchmarks should be done on non-system drive
While I can understand the logic, for the vast majority of the SSD reviews I have done I disagree.
The reasoning is pretty simple; if someone buys and uses the SSD I am reviewing, they likely won't operate it under 'benchmark conditions', so I try to run my benchmarks with the drive 'positioned' in a similar position as an end user. This means booting from it, and it means operating with the drive partially filled, and 'dirty' from a TRIM perspective.
Secondary to that, on many systems performance on a secondary drive can be impacted by the primary drive. Windows will perform virtual memory operations on the primary drive, even if all you are doing is benchmarking a secondary disk. In addition, I specifically install drives in M2_1 on my Ryzen system because that is the only M.2 slot that is CPU-attached; M2_2 is attached to the chipset PCIe lanes.
Specific to the P5800X on my Ryzen platform, I tend to agree with your statement a bit more; the P5800X is a specialized device and very lightly might be used as an 'accelerator' for a specific task, and not necessarily a boot drive. On the other hand, I wanted the results to be comparable to my previous benchmarks, so I elected to test it in the same manner as other SSDs I have reviewed.
Of course, I did run into problems. Since the P5800X is on a PCIe card rather than M.2, my first inclination was to install the PCIe card into the first PCIe slot (where the GPU would normally go) because this forces it to run from the CPU lanes just like operating in M2_1. However, when attached in this manner the P5800X experienced intermittent freezing and benchmarks that went no higher than 200 MB/s; obviously something was wrong. When attached via the second PCIe slot (and thus running off the chipset) performance was much better, but still receiving results 1-2 GB/s slower than other reviewers were getting with equivalent hardware. I tested briefly on an 8-core Epyc system I had handy and the sequential results were back to 'normal' so the problem was specific to my test platform. This is why I ran on the i5-11600k system, it had better results than my Epyc system, though was still disadvantaged compared to the normal 5900X test bench.