We propose a power efficient non-orthogonal multiple access (NOMA) scheme suitable for Internet of Things (IoT) and device-to-device (D2D) communications. OFDM-based power-domain NOMA (PD-NOMA) is one of the promising approaches to achieve simultaneous connections of massive devices. Nevertheless, the problem of high peak-to-average power ratio (PAPR) is a major challenge, since power amplifier (PA) efficiency is a critical factor for IoT and D2D devices consisting of low-cost hardware components. This paper overcomes this problem by introducing time-frequency domain NOMA (TF-NOMA), where OFDM and DFT-spread OFDM (DFT-s-OFDM) are superposed over the time-frequency domain. By utilizing the low PAPR property achieved by DFT-precoding, TF-NOMA can efficiently reduce the PAPR of the downlink signal transmitted by the base station (BS) as well as the uplink signals of the devices located far from BS. As a result, the proposed TF-NOMA can operate with higher PA efficiency than the conventional OFDM-based PD-NOMA. Furthermore, the insertion of DFT-precoding with the aim of PAPR reduction also enhances the error propagation resistance due to the resulting Gaussian mixture signal property. Throughout the mathematical analysis and computer simulations, we demonstrate that TF-NOMA achieves better error rate performance than the conventional PD-NOMA with significant improvement of PA efficiency.