It's more than you'd think. It really depends on your standard deviation.

If you winrate is 17/hour, I'm guessing your std. deviation is in the ballpark of 60-70.

If this is the case, after 600 hours your estimate is within 5-6 of the true mean - your winrate is between $12-22 (a 95% confidence inteval)

How many hours you need for an "accurate winrate" is basically how accurate you need to be - after 1,000 hours you will be +/- 4 away

after 2,000 hours you will be +/- 2.5 away

If you want to really get this right, you can estimate your standard deviation (which is much easier to get a good estimate of than winrate) The confidence rate will be +/- (2*stddev/(hours^.5)) With your winrate, I'd be surprised if you std. deviation wasn't right around 60-75 though.