This paper describes a high-speed computational method of the fuzzy inference system (FIS) for embedded systems. The proposed method adopts the composition of the variable structure system (VSS). The proposed method was converted into the calculation process of the variable separation formation by making it to linear systems in each domain. It is shown that this proposed method is effective for the FIS in three inputs and over especially. The amount of stored data by the proposed method is about 98% less than the table look-up method (TLM) for the FIS in three inputs and 25 rules. Moreover, the computational time by the proposed method is about 90% less than a classical method by the same system