[jmalicki]

Speaking as someone who actually runs NVMe RAIDs for large-ish databases, the NVMe speeds quoted are usually only attainable for large sequential streaming loads. While they are not spinning disks with heads that need to seek, there is still very substantial random I/O overhead with SSDs, including NVMe. Even with a top of the line model like a Samsung 980 Pro, you might only get 5% the PCIe 4 bus speed with heavy random IO performance.

RAM has cachelines and such, but postgres reading 8k at a time isn't limited by that.

How to work something like this into Linux as say, a very fast swap device, is another question.

[namibj]

The 980 Pro does way more with random reads of sufficient concurrency. The main issue is the 100us latency under load, so to get half a million IOPS, that's 50 concurrent requests. For maximum speed you need some batching I think.