[Paper Review] HDTorch: Accelerating Hyperdimensional Computing with GP-GPUs for Design Space Exploration

William Simon, Una Pale+2 more|arXiv (Cornell University)|Jun 9, 2022

Ferroelectric and Negative Capacitance Devices35 references4 citations

TL;DR

HDTorch is an open-source, PyTorch-based library with CUDA-accelerated hypervector operations for Hyperdimensional Computing (HDC), enabling fast design space exploration. It achieves up to 111x speedup in classical HDC training and 87x in inference, while enabling the first full-dataset analysis of the 980-hour CHB-MIT EEG epilepsy dataset, revealing critical limitations of subset-based training.

ABSTRACT

HyperDimensional Computing (HDC) as a machine learning paradigm is highly interesting for applications involving continuous, semi-supervised learning for long-term monitoring. However, its accuracy is not yet on par with other Machine Learning (ML) approaches. Frameworks enabling fast design space exploration to find practical algorithms are necessary to make HD computing competitive with other ML techniques. To this end, we introduce HDTorch, an open-source, PyTorch-based HDC library with CUDA extensions for hypervector operations. We demonstrate HDTorch's utility by analyzing four HDC benchmark datasets in terms of accuracy, runtime, and memory consumption, utilizing both classical and online HD training methodologies. We demonstrate average (training)/inference speedups of (111x/68x)/87x for classical/online HD, respectively. Moreover, we analyze the effects of varying hyperparameters on runtime and accuracy. Finally, we demonstrate how HDTorch enables exploration of HDC strategies applied to large, real-world datasets. We perform the first-ever HD training and inference analysis of the entirety of the CHB-MIT EEG epilepsy database. Results show that the typical approach of training on a subset of the data does not necessarily generalize to the entire dataset, an important factor when developing future HD models for medical wearable devices.

Motivation & Objective

• To address the lack of efficient, flexible, and open-source frameworks for Hyperdimensional Computing (HDC) that support rapid design space exploration.
• To accelerate HDC training and inference using GPU-accelerated CUDA kernels for hypervector operations such as bit-packing and summation.
• To evaluate HDC performance on large, real-world biomedical datasets—particularly the unbalanced CHB-MIT EEG database—where prior work relied on data subsets.
• To investigate the impact of hyperparameters like hypervector width and online training batch size on accuracy and runtime.
• To demonstrate that subset-based training may not generalize to full datasets, especially in highly imbalanced medical applications.

Proposed method

• HDTorch extends PyTorch with custom CUDA kernels for efficient hypervector operations, including horizontal and vertical bit-array summation and bit-(un)packing.
• It supports both classical and online HDC training, with online learning enabling incremental updates using batched data to reduce computational overhead.
• The framework integrates with PyTorch’s autograd system, enabling end-to-end differentiable HDC pipelines for model optimization.
• Hypervector operations are optimized using parallel GPU execution, reducing memory consumption by up to 10x compared to CPU-based implementations.
• The system supports customizable encoding and similarity strategies, enabling flexible exploration of HDC design space.
• HDTorch is evaluated on four benchmark datasets and applied to the full CHB-MIT EEG dataset, comparing performance across data organization strategies (Fact10 vs. 1HSeg).

Experimental results

Research questions

• RQ1 Can GPU-accelerated hypervector operations significantly reduce training and inference time in HDC without sacrificing accuracy?
• RQ2 How do hyperparameters such as hypervector width and online training batch size affect the trade-off between accuracy and runtime in HDC?
• RQ3 Does training on a data subset generalize to performance on the full, large-scale, unbalanced CHB-MIT EEG dataset?
• RQ4 Can HDTorch enable the first full-dataset analysis of the CHB-MIT epilepsy database, revealing insights not visible in prior subset-based studies?
• RQ5 How does the performance of HDC compare to traditional models like random forests when applied to the entire CHB-MIT dataset?

Key findings

• HDTorch achieves an average of 111x speedup in classical HDC training and 68x in online HDC training, with 87x speedup in inference across benchmark datasets.
• The use of CUDA-optimized kernels reduces memory consumption for hypervector operations by up to 10x compared to CPU-based implementations.
• Batching in online HDC training reduces training time significantly—up to 68x speedup—but causes an average 7% drop in accuracy at large batch sizes (e.g., 8192).
• Performance on the full CHB-MIT dataset differs markedly from subset-based training: classical HDC shows a significant drop in F1 score and increased false positives when trained on the entire dataset.
• Random forest and online HDC outperform classical HDC on the full dataset, with online HDC avoiding the precision degradation seen in classical HDC, suggesting that training strategy matters for large, unbalanced datasets.
• The study reveals that typical subset-based training approaches may not generalize to full datasets, highlighting the need for full-dataset evaluation in medical HDC applications.