System-on-Chip (SoC) task allocation with memory size information can improve the data storage and transfer efficiency. Nowadays, the methods of memory size estimation often adopt the single granularity and run at slow speed. Therefore, the space exploration of multi-granularity task allocation cannot be supported well. For SoC task allocation, this paper proposes a fast, multi-granularity approach of memory size estimation. The approach estimates the memory size at three granularities. The technique of data domain partition is introduced into the time-consuming process of memory size estimation at loop granularity, which greatly accelerates the analysis processing.