Efficient GRAPPA or SENSE reconstruction is largely dependent on coil geometry in the direction in which phase encoding steps reduction is performed during partially parallel acquisition. In this study we demonstrate the ability to perform a 2D CAIPIRINHA trajectory in a 3D EPI sequence to reduce the geometry factor (g-factor) noise amplification in the reconstructed images for a predefined total acceleration. 2D CAIPIRINHA style k-space patterns provide improved reconstructions when using very large accelerations on one phase-encode direction, thanks to the ability to use the coil sensitivities along the other phase-encode direction to compensate the reduced coil sensitivity variation.