The mechanisms underlying cardiac fibrillation have been investigated for over a century, but we are still finding surprising results that change our view of this phenomenon. The present study focuses on the transition from normal rhythm to atrial fibrillation associated with a gradual increase in the pacing rate. While some of our findings are consistent with existing experimental, numerical, and theoretical studies of this problem, one result appears to contradict the accepted picture. Specifically we show that, in a two-dimensional model of paced homogeneous atrial tissue, transition from discordant alternans to conduction block, wave breakup, reentry, and spiral wave chaos is associated with transient growth of finite amplitude disturbances rather than a conventional instability. It is mathematically very similar to subcritical, or bypass, transition from laminar fluid flow to turbulence, which allows many of the tools developed in the context of fluid turbulence to be used for improving our understanding of cardiac arrhythmias.Atrial fibrillation is a common type of cardiac arrhythmia characterized by spatially uncoordinated high-frequency patterns of electrical excitation waves. Several different explanations of how fibrillation arises from a highly coordinated and regular normal rhythm have been offered. The most common is a dynamical scenario that involves several stages. First, a milder type of arrhythmia known as alternans, characterized by regular spatial and temporal modulation in the duration of the excitation, is created as a result of a period doubling or Hopf bifurcation. In the second stage, the modulation grows until some portion of an excitation wave fails to propagate (which is known as conduction block), producing wave breakup. In later stages, spiral waves segments, or wavelets, form and undergo subsequent breakups until a complicated dynamical equilibrium is established, with wavelets continually annihilating and reemerging. The current paradigm is that each of these stages represents a separate bifurcation that leads to either a change in the stability, or the disappearance, of the solution describing a previous stage. Our analysis of a simple model of paced cardiac tissue shows that transitions between different stages can happen without any bifurcations taking place. In particular, conduction block can be caused by strong transient amplification of finite amplitude disturbances, which is a flip side of a very well-known memory effect that describes the sensitivity of the asymptotic dynamics of paced cardiac tissue to the entire pacing history.