Lawrence Berkeley National Laboratory

Data movement remains a major bottleneck in high-performance computing workflows, making GPU-accelerated compression an increasingly important optimization. While NVIDIA platforms benefit from mature compression libraries such as nvCOMP, the performance of practical lossless compression on AMD GPUs remains undercharacterized. This paper addresses that gap by porting LZ4, Snappy, and Cascaded from CUDA to HIP and evaluating them across four AMD GPUs spanning three CDNA generations (MI50, MI210, MI300X) and RDNA 3 (RX 7900 XT), using both synthetic datasets and seismic simulation data. Our study makes three contributions. First, we provide a cross-generational characterization of lossless GPU compression on recent AMD accelerators. Second, we present a functional CUDA-to-HIP port of three widely used algorithms, establishing an open baseline for future AMD-specific optimization. Third, we analyze how architectural differences shape compression and decompression behavior across algorithms and datasets. The results show that the MI300X achieves up to 11×$$11\times $$ higher decompression throughput than the MI50, while RDNA 3 is competitive for compression workloads characterized by irregular memory accesses. We also show that transfers dominate end-to-end execution time on all evaluated platforms, indicating that the main benefits of GPU compression are most likely to emerge in GPU-resident workflows. For realistic seismic floating-point data, lossless compression ratios remain modest (1.03-1.04×$$1.03-1.04\times $$), suggesting that error-bounded lossy compression is a promising direction for future work.