Abstract: Time-dependent, whole-core calculation with high-fidelity pin-resolved details serves an important role in multi-physics reactor applications. However, the limitation of the current CPU has become a sore point. In order to maintain the high accuracy and reduce the computational burden, it is necessary to develop the advanced technology to speed-up the transient calculation. Recently, heterogeneous computing has been increasingly widespread used with the high-performance computing (HPC) systems, which is widely equipped in HPC clusters. Compared to CPU, current GPU has much higher rate of FLOPs/s (floating-point operations per second), larger memory bandwidth, and lower energy consumption per FLOP. These features leverage the parallel compute engine in GPUs to solve many complex computational problems in a more efficient way. Method of Characteristics (MOC) has been widely used in GPU-based whole-core transport calculation for its feature of natural massive parallelization. Based on the above points, a GPU-based 2D MOC transient fixed source problem (TFSP) solver is implemented on lattice physics code ALPHA (advanced lattice physics code based on heterogeneous architecture). In this paper, the full implicit method (FIM) were adopted to solve the TFSP. The Jacobi transport sweep algorithm was introduced in MOC TFSP solver. The transport sweep over energy groups and polar angles were performed for each ray segment during the ray tracing. To accelerate the convergence of the TFSP, the pin-based coarse mesh finite difference (CMFD) was implemented in the TFSP solver. In order to parallelize the CMFD TFSP solver on GPU, the red-black ordering algorithm was introduced in the CMFD TFSP solver. The results of two problem are presented including the TWIGL benchmark problem and the 2D MINI-CORE benchmark problem to verify the capability of the GPU-based TFSP solver. Numerical results demonstrate that the TFSP solver built in ALPHA has the desired accuracy. The percent difference of core power history between ALPHA and the reference are less than 0.5%. Compared with serial CPU-based solver, the speedup of the GPU-based solver is between 2.0x to 6.0x and the speedup is in inverse proportion to the time ratio of the CMFD calculation.