GPU-accelerated single-cell analysis at scale with rapids-singlecell

Abstract: Single-cell sequencing technologies reveal cellular heterogeneity at high resolution, advancing our understanding of biological complexity. As datasets start to scale to tens of millions of cells, computational workflows face substantial bottlenecks, with CPU-based analytical pipelines requiring hours or days for routine processing steps like filtering, normalization, and clustering. These scalability limitations fundamentally restrict common interactive data exploration and iterative hypothesis testing. Here we introduce rapids-singlecell, a GPU-accelerated framework that integrates natively with the scverse ecosystem and operates directly on the AnnData data structure, which delivers orders-of-magnitude speedups for single-cell workflows. Built on CuPy arrays and the NVIDIA CUDA-X Data Science (RAPIDS) ecosystem, rapids-singlecell provides near drop-in GPU replacements for core scanpy-based analysis steps. Across standard single-cell workflows such as preprocessing, dimensionality reduction, neighborhood graph construction, clustering, and batch correction, rapids-singlecell achieves speedups of up to several hundred-fold compared to optimized CPU baselines. This reduces analysis time from hours to minutes on standard hardware, while maintaining consistent biological interpretations. These performance improvements make it possible to analyze large data sets in close to real time, without the need for data splitting. Together with real-time parameter tuning and iterative workflows, rapids-singlecell makes interactive large-scale single-cell analysis possible.

• The paper presents rapids-singlecell, a GPU-accelerated framework that reduces single-cell analysis runtime by over 120-fold compared to traditional CPU workflows.
• It employs custom CUDA-optimized kernels and integrates seamlessly with libraries like cuML and cuGraph for efficient sparse matrix operations and scalable analytics.
• Benchmark results demonstrate dramatic performance gains, enabling interactive analysis of millions of cells in seconds to minutes, facilitating rapid hypothesis testing and data integration.

GPU-Accelerated Single-Cell Analysis: The rapids-singlecell Framework

Motivation and Context

Modern single-cell sequencing technologies have enabled unprecedented views into cellular heterogeneity and complexity, driving fundamental biological insights. However, as datasets routinely surpass millions of cells, established CPU-based computational workflows become prohibitively slow, impeding interactive analyses, hypothesis iteration, and large-scale data integration. Bottlenecks arise from highly parallelizable operations—matrix transformation, nearest-neighbor search, clustering—but traditional pipelines fail to leverage GPU architectures optimized for these tasks. Existing toolchains lack transparent integration of GPU acceleration, creating friction for adoption at-scale in single-cell studies.

Framework Architecture and Technical Implementation

rapids-singlecell addresses these bottlenecks by implementing a tightly integrated, GPU-accelerated analytical pipeline for single-cell omics, compatible with the prevailing scverse ecosystem and the AnnData data structure. The architecture leverages CuPy for sparse matrix operations, the RAPIDS ecosystem (cuML for machine learning, cuGraph for graph analytics, cuVS for vector search), and Dask for parallel out-of-core and multi-GPU execution. All core steps—normalization, highly variable gene selection, dimensionality reduction, neighbor graph construction, clustering, and batch correction—are executed with custom CUDA-optimized kernels, achieving high throughput and efficient memory utilization. Memory management uses the RAPIDS Memory Manager (RMM) with oversubscription, allowing analyses beyond local device memory by spilling to host RAM as needed.

The sparse matrix primitives are critical: gene expression matrices are highly sparse, and rapids-singlecell avoids unnecessary densification or host-device transfer. Kernel design is tuned for cell-wise and gene-wise operations (CSR and CSC formats), with lazy format conversion and caching for minimal overhead. The framework mirrors scanpy’s API, enabling near drop-in replacement of CPU routines; users can migrate to GPU-based pipelines with minimal code changes. For stochastic and non-deterministic algorithms (PCA, UMAP, clustering), output equivalence is validated at the biological interpretation level, and deterministic operations are numerically consistent within floating-point tolerance.

Performance Characterization

Benchmarks using 10x Genomics and other atlas-scale datasets demonstrate rapids-singlecell’s substantial speedups. For a 1.3 million cell mouse brain dataset, the end-to-end workflow is reduced from >52 minutes on a 32-core workstation to 26 seconds on a single NVIDIA DGX B200 GPU, a >120-fold improvement. Individual steps see core gains: preprocessing (~70×), UMAP (~350×), Leiden clustering (~100×). Speedup is consistent across architectures: RTX PRO 6000 (Blackwell WS) and H200 GPUs yield ~120× and ~110× speedups, respectively, with pooled memory allocation.

Batch correction using the Harmony algorithm, GPU-optimized within rapids-singlecell, achieves ~100× speedup on 200,000-cell datasets and >250× for 2.4 million cells, while processing an 11.4 million cell dataset in under 25 seconds—where CPU baselines fail within a two-hour window. With Dask-based chunking and multi-GPU execution, the Tahoe-100M dataset (100 million cells) is analyzed in under 20 minutes on an 8-GPU node. For smaller datasets (<50,000 cells), kernel launch overhead and memory transfer costs can limit speedup, illustrating the framework’s design bias toward large-scale use cases.

Validation and Reproducibility

Outputs for deterministic operations are numerically validated against scanpy baselines (e.g., normalization, HVG selection, PCA), while stochastic algorithms are evaluated for consistent biological interpretation (cluster assignments, neighborhood structure preservation). Minor numerical differences are attributed to GPU floating-point accumulation order and atomic operations, a well-documented phenomenon in GPU computing. The Harmony implementation maintains a Pearson correlation >95% for corrected principal components versus reference baselines. Users requiring bitwise reproducibility for regulatory purposes should document hardware and software versions explicitly.

Practical Implications and Theoretical Impact

rapids-singlecell fundamentally transforms the analytical paradigm for single-cell studies. Dramatic reductions in runtime enable interactive parameter sweeps, real-time exploratory analysis, rapid embedding comparisons, and iterative hypothesis testing previously infeasible in clinical or translational contexts. Data splitting, batching, and delayed evaluation—as seen in CPU-bound workflows—are eliminated. For atlas-scale and spatial transcriptomics datasets, this shift supports efficient construction, integration, and interrogation of data spanning tens to hundreds of millions of cells.

The technical innovations—transparent GPU integration with community-standard data structures and APIs, modular hardware-agnostic scaling, and custom kernel optimization—lower barriers to adoption and future method development. By maintaining compatibility with scanpy and scverse, rapids-singlecell can serve as a foundation for integrating emerging single-cell analytic methods, including spatial transcriptomics and foundational cell biology models.

Limitations and Future Directions

Despite major gains for large datasets, rapids-singlecell’s speedup is diminished for small datasets due to fixed GPU overhead costs. Memory-constrained consumer GPUs (<8 GB VRAM) preserve functionality at reduced performance when oversubscribing to host RAM. Mixed workflows with unimplemented steps require CPU fallback and incur data transfer overhead. Current framework is restricted to CUDA-capable hardware, though broad device support is planned.

Planned extensions include additional support for spatial transcriptomics workflows, integration with cell biology foundation models, and further hardware-agnostic performance optimizations. Automatic delegation from scanpy to rapids-singlecell is under development for seamless GPU acceleration within established pipelines.

Conclusion

rapids-singlecell delivers robust, scalable, and modular GPU-accelerated single-cell analysis. By enabling order-of-magnitude reductions in computation time without sacrificing biological interpretability, the framework offers a transformative acceleration of routine single-cell workflows, supporting interactive and iterative analytics at the scale required for modern omics studies. Its architecture and compatibility foster adoption alongside rapid incorporation of future methods, positioning GPU acceleration as a critical infrastructure for single-cell and broader omics domains.