The 3D graphics is becoming an important part in modern multimedia systems from the architectural design to the desktop games. The 3D graphics pipeline which processes 3D objects and show these images to the 2D screen requires a lot of floating-point and integer operations. The 3D pipeline is composed of 3 stages : i.e, scene modeling, the geometry stage, and the rendering stage. Generally, the geometry stage requires a large number of floating-point operations and is handled by CPU. The rendering stage requires lots of integer operations and is processed at the graphics card which has on-board 3D accelerator such as Voodoo2, RIVA, and i740. These 3D accelerators for rendering stage has been researched and their performances are about 100M pixel/s which is sufficient for real-time 3D image. The bottleneck of the 3D pipeline becomes the geometry stage which is responsible for the transformation, clipping, and lighting of the objects. This thesis presents a new floating-point unit which is designed for 3D graphics geometry engine. It is used in FLOVA which is VLIW architecture. It has fastpow unit which computes the power operation for the lighting of the vertex. The fast vector normalization using Newton-Raphson method and the efficient scheduling of 4x4 matrix multiplication for transformation and clipping are supported in the designed floating-point units.