SQL Server license disallows publishing of the results of benchmarking (much like Oracle does)

Wait, really? That's kind of messed up.

Remarkable that no throwaway HN accounts considered that a challenge.

Likely very similar to the postgres benchmarks. Fundamentally a RDBMS needs to sync each transaction commit to the log file on disk and that sync is always a syscall. If your DB is doing thousands tx/sec to low latency flash and you rely on that low latency, you're going to get hit.

Note that the postgres benchmark numbers passed around (most based on my benchmarks) are readonly. For write based workloads the overhead is likely going to be much smaller in nearly all cases, there's just more independent workload. The overhead in the readonly load profiles comes near entirely from the synchronous client<->server communication, if you can avoid that for workloads (using pipelining, other batching techniques) the overhead's going to be smaller.