In this paper, a high throughput multiple transform architecture for H.264 Fidelity Range Extensions (FRExt) is proposed. New techniques are adopted which (1) regularize the $8{ imes}8$ integer forward and inverse DCT transform matrices, (2) divide them into four $4{ imes}4$ sub-matrices so that simple fast butterfly algorithm can be used, (3) because of the similarity of the sub-matrices, mixed butterflies are proposed that all the sub-matrices of $8{ imes}8$ and matrices of $4{ imes}4$ forward DCT (FDCT), inverse DCT (IDCT) and Hadamard transform can be merged together. Based on these techniques, a hardware architecture is realized which can achieve throughput of 1.488Gpixel/s when processing either $4{ imes}4;or;8{ imes}8$ transform. With such high throughput, the design can satisfy the critical requirement of the real-time multi-transform processing of High Definition (HD) applications such as High Definition DVD (HD-DVD) ($1920{ imes}1080@60Hz$) in H.264/AVC FRExt. This work has been synthesized using Rohm 0.18um library. The design can work on a frequency of 93MHz and throughput of 1.488Gpixel/s with a cost of 56440 gates.