We've got ours set to 90 on a forum with 8 CPUs (I think that's 2 chips x 2 cores x 2 hyperthreading) and maybe 800 concurrent users max, 250 average.
We get Server Load exceeded messages more often than I would expect, when the server isn't being hit by peak load. If I look at perfmon (we're running on windows2008) then I see a lot of short spikes, so something hits the CPU hard in short bursts, but an average of something like 30-40% most of the time across all cores.
So is this setting asking windows what the average %CPU load is over a period of time? Or Instantaneous? Does it respect multiple cores? I guess it has to these days? Is it a percentage? I'm guessing that it is, but it doesn't say in the ACP.