A Dependency-Aware Multi-GPU Training Framework for Temporal Graph Neural Networks

Name: A Dependency-Aware Multi-GPU Training Framework for Temporal Graph Neural Networks
Uploaded: 2026-10-18T19:39:04.575972+00:00
Duration: 397 s
Description: DepBloom accelerates multi-GPU Temporal Graph Neural Network (TGNN) training by integrating three techniques: temporally bounded spatial partitioning, intra-batch memory refinement, and critical-path optimization. This framework handles complex dependencies more flexibly than existing methods, leading to significant speedups (up to 5.16x) while maintaining high model quality.

DepBloom accelerates multi-GPU Temporal Graph Neural Network (TGNN) training by integrating three techniques: temporally bounded spatial partitioning, intra-batch memory refinement, and critical-path optimization. This framework handles complex dependencies more flexibly than existing methods, leading to significant speedups (up to 5.16x) while maintaining high model quality.

Open the full episode