ResearchBenchmarksReduction-based convolutions on tolokers 2Reduction-based convolutions on tolokers 2Follow2.54Latency Speedup (Forward Pass)Ours0.3560.9231.492.057May 29, 2026Evaluation resultsMethodMethodLinksLatency Speedup (Forward Pass)Latency Speedup (Backward Pass)Peak Memory Ratio (Forward Pass)Peak Memory Ratio (Backward Pass)OursHidden dimension (D)=5...Hidden dimension (D)=512, Aggregation operator=min-aggregation2026.052.541.321.971.51cuGraphHidden dimension (D)=5...Hidden dimension (D)=512, Aggregation operator=min-aggregation2026.050.440.871.191.12