Verified, then logged.
Sun 22 Nov 2026 · 00:33 KST / 15:33 UTC한국어 | ENSearch /

Breaking02 / IT

NVIDIA ties Ling-3.0-flash bottleneck on DGX Spark to memory

NVIDIA said reading model weights drove execution time in a batch-size-1 evaluation of the Inclusion AI model.

A close-up of a palm-sized desktop AI computer with its lid removed on a workbench, measurement probes clipped to the memory chips surrounding its processor. AI-generated image

NVIDIA analysed the bottleneck in Ling-3.0-flash on DGX Spark in a blog post on 28 Oct 2026. The analysis is relevant to users of mixture-of-experts (MoE) models, which use only part of the neural network for each input.

NVIDIA said the evaluation ran the Inclusion AI model at a batch size of 1. Profiling, which measures the time spent in each part of a run, showed that decoding was memory-bound. The term means performance is limited by memory access rather than computation. Reading model weights drove execution time.

The company named the MoE computation and the output-layer projection as the operations that stood out. The output-layer projection is the step that produces the model's output scores.

The blog opens with the sentence "Generating tokens faster starts with identifying what limits inference performance." Its headline offers three ways to boost MoE inference.

Test conditions. The evaluation described in the blog is a run of Ling-3.0-flash on DGX Spark at batch size 1.