Compression is widely used to reduce storage costs and improve data transfer speeds across databases, data centers, high-performance computing, and deep learning. However, decompressing data often introduces latency and consumes valuable compute resources, creating performance bottlenecks. To solve this, NVIDIA has introduced the hardware Decompression Engine (DE) in the Blackwell architecture, paired with the nvCOMP library, enabling fast, efficient decompression directly in hardware. The Blackwell DE is a fixed-function unit designed to accelerate decompression of Snappy, LZ4, and Deflate-based data streams. By offloading decompression from the GPU’s streaming multiprocessors (SMs), the DE frees up compute capacity for actual workloads, such as AI training or scientific simulations. The DE is integrated into the GPU’s copy engine, allowing compressed data to be transferred over PCIe or C2C links and decompressed in transit—eliminating the need for separate host-to-device copies followed by software decompression. This design enables true concurrency: while data is being decompressed, the GPU can simultaneously execute compute kernels. This is especially beneficial for multi-stream workloads, where multiple data streams can be processed in parallel, keeping the GPU fully utilized and avoiding I/O bottlenecks. The nvCOMP library provides GPU-accelerated compression and decompression routines for a variety of standard and optimized formats. While CPUs and fixed-function hardware often outperform GPUs on standard formats due to parallelism limitations, the DE bridges this gap. Developers can use nvCOMP APIs to access the DE seamlessly. The library automatically detects whether the DE is available and uses it when possible, falling back to SM-based decompression otherwise—ensuring code portability across GPU generations. To use the DE, developers must allocate buffers with specific requirements. The DE supports allocations created via cudaMallocFromPoolAsync or cuMemCreate with the cudaMemPoolCreateUsageHwDecompress or CU_MEM_CREATE_USAGE_HW_DECOMPRESS flags. These allocations must be pinned host memory and properly aligned. Using the same allocation for multiple buffers in a batch improves performance by reducing host driver overhead. If buffers come from different allocations, the overhead can become significant. Note that on B200 GPUs, any buffer larger than 4 MB will trigger a fallback to SM-based decompression. This limit may change over time and can be queried programmatically. In terms of performance, the DE offers significantly faster decompression than SM-based approaches for most workloads, especially at smaller chunk sizes. The DE has dozens of execution units, each optimized for decompression tasks, while SMs offer thousands of warps but are less efficient for this specific function. For certain workloads, SMs can match DE performance when fully saturated, but the DE still provides better overall system efficiency by freeing up SMs. Performance comparisons using the Silesia benchmark show that the DE outperforms SM-based decompression for Snappy, LZ4, and Deflate at 64 KiB and 512 KiB chunk sizes, with further optimization potential in future software updates. To get started, developers should use nvCOMP APIs with properly configured allocations. The library handles the complexity of hardware detection and fallback, enabling faster, more efficient data pipelines with minimal code changes. This makes it easier than ever to scale data-intensive applications on NVIDIA Blackwell GPUs.