Floating-point multipliers are typically power hungry which is undesirable in many embedded applications. This paper proposes a variable-latency floating-point multiplier architecture, which is suitable for low-power, high-performance, and high-accuracy applications. The architecture splits the significand multiplier into upper and lower parts, and predicts the required significand product and sticky bit from upper part. In the case of correct prediction, the computation of lower part is disabled and the rounding operation is significantly simplified so that floating-point multiplication can be completed early. Finally, detailed design and simulation of the floating-point multiplier is presented, together with its evaluation by comparing power consumption with the fast and conventional floating-point multipliers. Experimental results demonstrate that the proposed double-precision multiplier consumes up to 26.41% and 24.97% less power and energy than the fast floating-point multiplier respectively at the expense of only small area and delay overhead. In addition, the results also show that the performance of proposed floating-point multiplier is very approximate to that of fast floating-point multipliers.