Since the size of time-varying volumetric data sets typically exceeds the amount of available GPU and main memory, out-of-core streaming techniques are required to support interactive rendering. To deal with the performance bottlenecks of hard-disk transfer rate and graphics bus bandwidth, we present a hybrid CPU/GPU scheme for lossless compression and data streaming that combines a temporal prediction model, which allows to exploit coherence between time steps, and variable-length coding with a fast block compression algorithm. This combination becomes possible by exploiting the CUDA computing architecture for unpacking and assembling data packets on the GPU. The system allows near-interactive performance even for rendering large real-world data sets with a low signal-to-noise-ratio, while not degrading image quality. It uses standard volume raycasting and can be easily combined with existing acceleration methods and advanced visualization techniques.