We empirically assess the relationship between population growth rate (lambda, a parameter central to ecology) and effective population size (N(e), a key parameter in evolutionary biology). Recent theoretical and numerical studies indicate that in semelparous species with variable age at maturity (such as Pacific salmon, many monocarpic plants, and various other species), differences in mean reproductive success among individuals reproducing in different years leads to variation in lambda, and this in turn can reduce N(e). However, this phenomenon has received little empirical evaluation. We examined time series of abundance data for 56 populations of chinook salmon (Onchorhynchus tshawytscha) from the northwestern United States and compared N(e) (calculated from demographic data) with the total number of spawners each generation (NT). Important results include: (1) The mean multigenerational ratio N(e)/N(T) was 0.64 (median = 0.67), indicating that annual variation in lambda reduces effective population size in chinook salmon by an average of approximately 35%. These reductions are independent of, and in addition to, factors that reduce N(e) within individual cohorts (uneven sex ratio and greater-than-random variance in reproductive success). (2) The coefficient of variation of lambda was the most important factor associated with reductions in N(e), explaining up to two-thirds of the variance in N(e)/N(T). (3) Within individual generations, N(e) was lower when there was a negative correlation between annual N(i) and lambda, i.e., when relatively few breeders produced relatively high numbers of offspring. Our results thus highlight an important and little-studied eco-evolutionary trade-off: density-dependent compensation has generally favorable ecological consequences (promoting stability and long-term viability) but incurs an evolutionary cost (reducing N(e) because a few individuals make a disproportionate genetic contribution). (4) For chinook salmon, N(eH) (an estimator based on the harmonic mean number of breeders per year) is generally a good proxy for true N(e) and requires much less data to calculate.