NVIDIA Shows 3 Ways to Speed Up MoE Inference on DGX Spark, Lifting Ling-3.0-flash Throughput 42.5%

NVIDIA demonstrated three stacked optimizations that speed up MoE inference on a single DGX Spark, lifting GSM8K decode throughput on Ant Group inclusionAI's 124B Ling-3.0-flash by 42.5% with no detectable accuracy change.