SunSpider should be rebased to run at least twice as long SunSpider now runs in under 4 seconds on my (laptop) hardware. The test should be rebased to take at least twice that long to help reduce error rates. Yes, that means it would be *even slower* in other browsers, but my long term goal is to move all browsers to at least 10x their current JavaScript speed. Rebasing the test to take 2x as long is a step in that direction.
Error rates is not the sole motivator. But certainly with more signal, small amounts of noise in the testing environment matter less.
Perhaps the tests should just be run multiple times, and then compute mean and variance?
You can already do that with --runs=15 But that doesn't really address the precision question in my mind.