Vector-matrix-multiplication (VMM) accelerators have gained a lot of traction, especially due to the rise of convolutional neural networks (CNNs) and the desire to compute them on the edge. Besides the classical digital approach, analog computing has gone through a renaissance to push energy efficiency further. A more recent approach is called time-domain (TD) computing. In contrast to analog computing, TD computing permits easy technology as well as voltage scaling. As it has received limited research attention, it is not yet clear which scenarios are most suitable to be computed in the TD. In this work, we investigate these scenarios, focussing on energy efficiency considering approximative computations that preserve accuracy. Both goals are addressed by a novel efficiency metric, which is used to find a baseline design. We use SPICE simulation data which is fed into a python framework to evaluate how performance scales for VMM computation. We see that TD computing offers best energy efficiency for small to medium sized arrays. With throughput and silicon footprint we investigate two additional metrics, giving a holistic comparison.