NVIDIA Expands DGX Spark to 8-Way Clusters as RTX Inference Jumps 40%
Updated
Updated · Wccftech · Oct 21
NVIDIA Expands DGX Spark to 8-Way Clusters as RTX Inference Jumps 40%
3 articles · Updated · Wccftech · Oct 21
Summary
DGX Spark now supports manual 8-system clusters, doubling the previous 4-way limit and lifting pooled memory to 512 GB on 64 GB units and 1 TB on 128 GB systems.
That larger memory footprint is aimed at frontier-class local AI workloads, with NVIDIA saying 1 TB can run models such as GLM 5.3 and support longer contexts, concurrent agents and distributed development.
40% faster inference is the top performance gain in the update: RTX PRO 6000 Blackwell gets that uplift on Qwen 3.8 Flash-Next, while DGX Spark gains 30% on Qwen 3.8 27B and DGX Station rises 20% on GLM 5.3 Flash.
Later this month, NVIDIA Sync will add Model Launcher and 8-system support to simplify downloading models, connecting them to OpenCode and managing troubleshooting and maintenance through updated DGX Spark Skills.
The GTC Berlin rollout also adds new local models, RAM-efficient checkpoints and a Wi-Fi Assistant, ahead of DGX Spark 64 GB shipments starting Oct. 23.