--- format: "aidr-story-markdown/v1" id: "0e62b0d0f4210d10540fc805535d37fbeb18d87b1889b8cca8e803c5cd71faf8" canonical_url: "https://aidr.today/0e62b0d0?lang=en" title: "3 Ways to Boost MoE Inference on NVIDIA DGX Spark" lang: "en" requested_lang: "en" available_langs: ["en","vi"] translation_fallback: null fallback_fields: [] published_at: "2026-10-27T15:00:00.000Z" category: "Infra" topics: ["nvidia","inference"] source_urls: ["https://developer.nvidia.com/blog/3-ways-to-boost-moe-inference-on-nvidia-dgx-spark/"] summary: "Profiling of Ling-3.0-flash on NVIDIA DGX Spark found decoding to be memory-bound, with the mixture-of-experts computation and output-layer projection as main bottlenecks, leading to three techniques to reduce weight traffic and increase tokens per verification round. The model uses 124 billion total parameters with 5.1 billion active parameters per token across 42 layers, routing each token to eight of 512 experts plus one shared expert. The post explains each optimization, measures its effect, and notes an integration issue that can make the combined configuration slower than baseline." --- # 3 Ways to Boost MoE Inference on NVIDIA DGX Spark > [Open the canonical story]() **Published:** 2026-10-27T15:00:00.000Z **Category:** Infra **Topics:** nvidia, inference ## Summary Profiling of Ling\-3\.0\-flash on NVIDIA DGX Spark found decoding to be memory\-bound, with the mixture\-of\-experts computation and output\-layer projection as main bottlenecks, leading to three techniques to reduce weight traffic and increase tokens per verification round\. The model uses 124 billion total parameters with 5\.1 billion active parameters per token across 42 layers, routing each token to eight of 512 experts plus one shared expert\. The post explains each optimization, measures its effect, and notes an integration issue that can make the combined configuration slower than baseline\. ## Sources - [Story source]()