Scholarlink

Hidden Markov Models (HMMs), along with their extension, Hidden Semi-Markov Models (HSMMs), are powerful tools for modeling complex systems with multivariate states, but scaling them to ultra-large state spaces presents significant computational and memory challenges. This paper proposes a hybrid CUDA-based implementation to overcome these limitations, enabling efficient processing of HMMs with massive state spaces and extended observation sequences. Key optimizations include log-space computations for numerical stability, memory-efficient data partitioning with sparse matrix representations, and asynchronous data transfers using CUDA streams to overlap host-device communication with GPU kernel execution. Our approach achieves significant performance improvements, demonstrating up to \(8.5\times \) speedup over multi-threaded CPU implementations for HMM processing.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

GPU-Accelerated Out-of-Core HMM Inference with Concurrent CUDA Streams

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • MohammadReza HoseinyFarahabady,
  • Albert Y. Zomaya

摘要

Hidden Markov Models (HMMs), along with their extension, Hidden Semi-Markov Models (HSMMs), are powerful tools for modeling complex systems with multivariate states, but scaling them to ultra-large state spaces presents significant computational and memory challenges. This paper proposes a hybrid CUDA-based implementation to overcome these limitations, enabling efficient processing of HMMs with massive state spaces and extended observation sequences. Key optimizations include log-space computations for numerical stability, memory-efficient data partitioning with sparse matrix representations, and asynchronous data transfers using CUDA streams to overlap host-device communication with GPU kernel execution. Our approach achieves significant performance improvements, demonstrating up to \(8.5\times \) speedup over multi-threaded CPU implementations for HMM processing.