AMPED: Accelerating MTTKRP for Billion-Scale Sparse Tensor Decomposition on Multiple GPUs

2025-07-20

AI Summary

To address memory and throughput bottlenecks in Matricized Tensor Times Khatri–Rao Product (MTTKRP) computation for large-scale sparse tensor decomposition—particularly on billion-scale non-zero tensors—this paper proposes AMPED, a multi-GPU parallel algorithm. AMPED innovatively integrates fine-grained sparse tensor tiling, cross-GPU data partitioning, and dynamic load-balancing scheduling to significantly mitigate inter-GPU load imbalance and idle waiting. Evaluated on real-world billion-scale sparse tensors under a single-node 4-GPU configuration, AMPED achieves a 5.1× geometric mean speedup over the state-of-the-art baseline methods. It is the first approach enabling efficient, single-node decomposition of ultra-large-scale sparse tensors, thereby overcoming fundamental memory and computational limitations inherent to single-GPU systems.

Technology Category

• Machine Learning: Matrix & Tensor Methods
• Data Mining & Knowledge Management: Scalability, Parallel & Distributed Systems
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Security and Privacy: Privacy-enhancing technologies

Abstract

Matricized Tensor Times Khatri-Rao Product (MTTKRP) is the computational bottleneck in sparse tensor decomposition. As real-world sparse tensors grow to billions of nonzeros, they increasingly demand higher memory capacity and compute throughput from hardware accelerators. In this work, we present AMPED, a multi-GPU parallel algorithm designed to accelerate MTTKRP on billion-scale sparse tensors. AMPED scales beyond the limits of a single GPU, meeting both the memory and performance requirements of large-scale workloads. We introduce a partitioning strategy combined with a dynamic load balancing scheme to distribute computation and minimize GPU idle time. On real-world billion-scale tensors, AMPED achieves a 5.1x geometric mean speedup in total execution time over state-of-the-art GPU baselines using 4 GPUs on a single CPU node.

Problem

• Accelerating MTTKRP for billion-scale sparse tensors
• Scaling beyond single GPU memory and performance limits
• Reducing GPU idle time via dynamic load balancing

Innovation

• Multi-GPU parallel algorithm for MTTKRP acceleration
• Partitioning strategy with dynamic load balancing
• Scales beyond single GPU memory limits