Deploying resource-intensive Video LLMs often relies on Split DNNs, yet this introduces a severe transmission bottleneck-intermediate embeddings 1 can be$26 \times$larger than original videos. Lossy compression is an effective solution to the I/O bottleneck in machine learning [1]. However, no compression solution exists to address the high dimensionality and unique layout of these embeddings.

Next-generation sequencing (NGS) platforms generate multi-terabyte data, making compression a major bottleneck for storage, transfer, and downstream analysis.

Large-scale scientific applications generate massive floating-point data, making lossy compression an essential method to reduce storage needs and improve performance. However, variations in compression time for various reasons can negatively impact scheduling and workload balance. Existing empirical approaches lack accuracy and generalizability [1]. This work proposes a novel analytical method for accurately predicting the compression time of prediction-based lossy compressors. The method decomposes the compression process into four stages: prediction and quantization, frequency counting and codebook construction, Huffman encoding, and additional lossless encoding. By using the statistics method, we capture the causes of time variation and eliminates application-specific dependencies. We design a surrogate model to predict the time cost: using sampling and offline testing to get the key parameters, then use regression and some probability models to make the prediction.