We propose a multistart CMA-ES with equal budgets for two interlaced restart strategies, one with an increasing population size and one with varying small population sizes. This BI-population CMA-ES is benchmarked on the BBOB-2009 noiseless function testbed and could solve 23, 22 and 20 functions out of 24 in search space dimensions 10, 20 and 40, respectively, within a budget of less than $10^6 D$ function evaluations per trial.