With the increasing diversification of future-oriented tasks and the increasing complexity of test environment, the traditional pulse code modulation/frequency modulation (PCM/FM) telemetry system based on FPGA has some shortcomings, such as the tight coupling of hardware and software, the long development cycle, and the difficulty of upgrade and maintenance. It can not meet the needs of large-scale deployment of future telemetry systems for elasticity, flexibility and agility. In order to solve the above problems, a GPU-based non-coherent demodulation method for PCM/FM signals is designed. The parallel demodulation method based on GPU is designed according to the algorithm principle and GPU parallel method and the parallel implementation method of specific modules is analyzed from the 3 parts of parallel digital down conversion, parallel FIR filtering and parallel differential frequency discrimination. The experimental results show that the non-coherent frequency discrimination process based on GPU consumes about 426. 867 ms when the code rate is 10 Mbps and the data is processed every second, which meets the real-time requirements of demodulation, the demodulation gain is slightly better than the traditional PCM/FM telemetry baseband based on FPGA, with about 0.1 dB demodulation gain.