New Method Makes Multiple GPUs Work 3x Faster Together
Faster AI training and route-finding could mean cheaper, quicker results.
Deep Dive
A new paper introduces a multi-GPU concurrent queue system built with NVIDIA's NVSHMEM, using the Bellman-Ford algorithm as a case study. On four NVIDIA A100 GPUs, the queue achieved up to 3.92x speedup, with an average of 3.04x, over a single-GPU baseline. When applied to Bellman-Ford shortest-path processing, it reached up to 3.03x speedup, averaging 2.65x, across 10 graphs from the SuiteSparse Matrix Collection.
Key Points
- The method speeds up multi-GPU teamwork by up to 3.92x compared to one GPU.
- It's the first known multi-GPU version of the Bellman-Ford algorithm (used for routing and maps).
- Tested on 10 real-world network datasets, averaging 2.65x faster — so it works outside a lab.
Why It Matters
Cheaper and faster GPU teamwork means better AI, quicker maps, and lower costs for everyone.