>>All that is for a physical computer and CPU. Now imagine (maybe you don't have to ;)) that your Windows computers are virtualized, so there's the hypervisor and underlying host OS adding further complication.
>>
>>Benchmarking these days is really hard.
>
>On the contrary... they are quite soft and malleable, as you have just shown :).
>
>In that regard, I'm thinking - what sense does then the question "peak CPU" mean? Any app can peak at 100% at some time. So I'd take it to mean "peak CPU as shown by task manager". Or else the question is pointless.

For a shared Citrix server I probably should have phrased it as average percent of a CPU core. If while the app is being run it averages 5% CPU (waiting for slow disk and even slower human input) then one core could theoretically support 20 simultaneous users.

OTOH there may be times when one user hogs a core for extended periods, maybe running SELECTs during report generation etc. With 20 users on a core the other 19 might get starved. The OS and its task scheduler will probably try to mitigate that, switch other users over to less-busy cores etc. So if you want consistently good responsiveness you probably wouldn't want to size the server for 20 users per CPU core, you'd probably want fewer than that.

That's why they pay the infrastructure guys minimum wage ;)