Ilmu Komputer & AI editorial
Open AccessOA2026
FastPair: GPU-Optimized String Decoding
Reorganizing dictionary lookups for contiguous memory writes achieves 2.4–4.2× speedup over NVIDIA's Decompression Engine
Joseph Isaacs; Francesco Gargiulo; Peter Boncz; Robert Kruszewski; Nicholas Gates; Rossano Venturini; Will Manning; Martin Prammer· 2026· DOI 10.48550/arXiv.2609.15034
The core problem
Modern data systems compress data at rest and decompress it only when needed to preserve interconnect bandwidth. This design is often inefficient on GPU-based compute platforms because many conventional compression techniques exhibit serial data dependencies that limit GPU parallelism, leaving resources idle. Recent NVIDIA GPUs address this decoding deficiency through the Decompression Engine (DE), an on-die, fixed-function decompression accelerator for general-purpose compression formats such as Deflate, LZ4, and Snappy. Recent work has proposed string codecs that replace frequent substrings with fixed-width codes from a small, trained dictionary, making each code's lookup independent. While these lookups can run in parallel, the resulting scattered reads and short output writes still do not align well with GPU hardware, which handles contiguous memory accesses more efficiently. The authors—Joseph Isaacs, Francesco Gargiulo, Peter Boncz, Robert Kruszewski, Nicholas Gates, Rossano Venturini, Will Manning, and Martin Prammer—present FastPair to bridge this gap.
Innovation
On a B300 GPU, FastPair decodes ten real-world columns 2.4 to 4.2× faster than the Decompression Engine, reaching up to 1.6 TB/s. This speedup is consistent across the evaluated columns, demonstrating the effectiveness of the reorganization strategy. The achieved throughput of 1.6 TB/s highlights the potential of dictionary-based codecs when adapted to GPU memory access patterns.
Modern data systems compress data at rest and decompress it only when needed to preserve interconnect bandwidth. This design is often inefficient on GPU-based compute platforms because many conventional compression techniques exhibit serial data dependencies that limit GPU parallelism, leaving resources idle. Recent NVIDIA GPUs address this decoding deficiency through the Decompression Engine (DE), an on-die, fixed-function decompression accelerator for general-purpose compression formats such as Deflate, LZ4, and Snappy. Recent work has proposed string codecs that replace frequent substrings with fixed-width codes from a small, trained dictionary, making each code's lookup independent. While these lookups can run in parallel, the resulting scattered reads and short output writes still do not align well with GPU hardware, which handles contiguous memory accesses more efficiently. The authors—Joseph Isaacs, Francesco Gargiulo, Peter Boncz, Robert Kruszewski, Nicholas Gates, Rossano Venturini, Will Manning, and Martin Prammer—present FastPair to bridge this gap.
FastPair optimizes the existing dictionary decoding process by reorganizing lookups and assembling decoded substrings for contiguous output writes. The core idea is to transform the scattered, short writes typical of dictionary-based decoding into coalesced, contiguous writes that GPUs handle efficiently. The decoding pipeline can be conceptualized as follows:
Why it matters
The results indicate that the primary bottleneck in GPU-accelerated dictionary decoding is not the lookup itself but the memory access pattern. By reorganizing lookups and assembling substrings for contiguous writes, FastPair aligns the decoding process with GPU hardware strengths. This approach outperforms the fixed-function Decompression Engine, which is optimized for general-purpose formats but may not exploit the independence of dictionary codes. The 2.4–4.2× speedup suggests that specialized codecs can significantly benefit from GPU-aware decoding strategies. Future work may explore integration with other compression formats and broader GPU architectures.
Who should read this
CS practitioners and researchers
Opening member content…