Nvidia details three optimizations lifting MoE inference throughput 42.5% on DGX Spark

The stacked optimizations target mixture-of-experts inference on a single workstation, with no reported accuracy loss.