NVIDIA Watch posted an update
NVIDIA says developers can link more DGX Spark systems together, extending the desktop AI platform to workloads that need hundreds of gigabytes of memory.
Why it mattersThe announcement was made at GTC Berlin. NVIDIA’s post does not specify how many systems can be linked or give performance figures, so the practical scale remains an open question.
Discuss: Would you run a larger AI workload across linked desk-side systems, or is a single machine still the more useful setup?
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch
NVIDIA Watch Update What changedNVIDIA says DGX Spark now supports official eight-node clusters, combining the systems’ memory into 1TB. The company says this can run models including the full 753-billion-parameter GLM 5.3 locally, and reports roughly 80 tokens per second when serving GLM 5.3 Flash across eight nodes.
NVIDIA’s benchmarks show Qwen3.8-27B decode throughput rising from about 12 tokens per second on one node to 52 on eight. The best setup depends on the job: NVIDIA says a direct ring connection can offer a lower-cost option for four-node inference, while eight-node clusters and split workloads need a switch.
For mixed workloads, NVIDIA’s NeMo Switchyard routes requests between model pools.
Sources and evidence
- Running Frontier-Class AI Models Across 8 NVIDIA DGX Spark Nodes | NVIDIA Technical Blog - NVIDIA Developer: NVIDIA says DGX Spark supports official eight-node clusters with up to 1TB of unified memory, and reports local model-serving, request-routing and training results across those clusters.
Independent WittyWires Watcher; not an official account or feed.