GPU-Optimized CUROCKET Boosts ROCKET Efficiency by Up to 11‑Fold per Watt

Global: GPU-Optimized CUROCKET Boosts ROCKET Efficiency by Up to 11‑Fold per Watt

A new algorithm called CUROCKET enables GPU acceleration of the ROCKET time‑series classification method, delivering up to 11 times higher computational efficiency per watt compared with the original CPU‑only implementation. The work was posted to arXiv on January 26, 2026, and targets the growing demand for faster, energy‑efficient analysis of time‑series data in fields such as finance, healthcare, and IoT.

Background on ROCKET

ROCKET (RandOm Convolutional KErnel Transform) was introduced in 2019 as a feature‑extraction technique for time‑series classification. It generates a large set of random convolutional kernels, applies them to input series, and feeds the resulting features into a linear classifier such as Ridge regression. At the time of its release, ROCKET matched the accuracy of leading state‑of‑the‑art methods while requiring far less computational resources.

Challenges of GPU Implementation

Although convolution operations are inherently parallelizable, ROCKET’s use of heterogeneous kernels makes conventional GPU convolution libraries inefficient. The varying lengths, dilations, and padding schemes of the kernels prevent straightforward batching, leading to underutilization of GPU cores and increased memory traffic.

Proposed Solution

The authors present CUROCKET, an algorithm that restructures the kernel‑application process to exploit GPU parallelism despite kernel heterogeneity. By grouping kernels with compatible parameters and employing custom CUDA kernels, CUROCKET achieves high occupancy and reduces data movement, thereby improving computational efficiency per watt.

Performance Results

Benchmarking on a standard NVIDIA GPU shows that CUROCKET attains up to 11× higher efficiency per watt than the original CPU implementation, while maintaining classification accuracy comparable to the baseline. The authors also report substantial reductions in wall‑clock time for large‑scale datasets.

Availability and Impact

The CUROCKET source code is publicly available on GitHub (https://github.com/oleeven/CUROCKET). By making GPU‑accelerated ROCKET accessible, the project may lower the barrier for deploying advanced time‑series classifiers in resource‑constrained environments and encourage further research into efficient random‑kernel methods.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.