NVIDIA expands DGX Spark clusters to eight nodes for local AI
NVIDIA said it will officially support eight-node DGX Spark clusters, extending the desktop AI system from a four-node ceiling to configurations with 1TB of unified memory and higher throughput for inference, orchestration and small-model training. The company made the announcement at GTC Berlin in a blog post published Oct. 22, 2026, and said support in NVIDIA Sync is due by the end of the year.
Memory scale and model support
According to NVIDIA, a single DGX Spark provides 128GB of unified memory, four nodes reach 512GB, and eight nodes reach 1TB. The company said that capacity lets developers load frontier-class models that would not fit on smaller configurations, including the full GLM 5.3 model from Z.ai, which NVIDIA described as a 753B parameter system. In an example test, NVIDIA said GLM 5.3 Flash ran across eight nodes at roughly 80 tokens per second, which it characterized as suitable for interactive local development.
NVIDIA also said the full GLM 5.3 checkpoint runs across the eight-node cluster in NVFP4 format, with the model measuring about 470GB in that precision. In the company’s table, prefill throughput for the full model was reported at 1,600 TPS at one, four and eight concurrent users, while decode throughput rose from 31 TPS at one user to 67 TPS at eight users. Those figures were presented by NVIDIA in the Oct. 22 post and reflect the company’s benchmark setup, including 8K input sequence length for prefill.
Cluster networking and collective communication
NVIDIA said DGX Spark nodes can be connected either through a switch or in a direct ring topology, and that the best choice depends on workload. For tensor-parallel inference, the company described a Spark-optimized collective backend called RoCEnante, which it said reduces latency for small-message, low-concurrency workloads by using an all-to-all exchange followed by local reduction. NVIDIA said the backend is merged into FlashInfer and falls back to NCCL if startup checks fail.
In the company’s comparison using Qwen3.8-27B-NVFP4, decode scaling at eight nodes reached 4.43x with RoCEnante on a switch topology versus 3.05x without it. NVIDIA said that represented a 45% improvement in scaling efficiency. The same table showed decode throughput rising from 11.86 TPS on one node to 52.58 TPS on eight nodes with the switch-backed optimization. NVIDIA said NCCL remains the supported default and is the better option as cluster size and concurrency increase.
Throughput, topology and multi-model routing
For a dense Qwen3.8-27B-NVFP4 benchmark without speculative decoding, NVIDIA said prefill throughput increased from 2,635 TPS on one node to 5,105 TPS on eight nodes in a switch configuration, while decode throughput rose from 11.86 TPS to 52.58 TPS. The company said a physical ring can perform well for a single model spread across all nodes, while a switch becomes necessary for split cluster allocations, any-to-any connectivity and workloads that depend on point-to-point communication.
NVIDIA also used the post to describe Switchyard, its multi-model orchestration approach. In the company’s prototype, a classifier routed simple requests to OpenBMB’s MiniCPM5 2B and harder requests to Qwen3.8 Flash Next. NVIDIA said the routed agent matched the quality of the Qwen-only configuration on a 92-case tool-use benchmark, scoring 91/100, while freeing capacity on the cluster. The company said that routing some traffic to the smaller model increased aggregate throughput by about 1.5x when half of requests went to Qwen and about 3.0x when a quarter of requests went to Qwen and the rest to MiniCPM.
Near-linear pre-training on eight nodes
NVIDIA said training scales differently from inference because each node holds a full copy of the model and only gradient synchronization crosses nodes once per step. To illustrate that, the company reported pre-training results for a nanochat model with about 896.5 million parameters, 20 transformer layers and hidden dimension sizes of 1,280 and 2,048-token sequences. Throughput increased from 8,741 TPS on one node to 68,079 TPS on eight nodes, which NVIDIA said equates to a 7.79x speedup and 97.4% scaling efficiency.
The company framed the result as evidence that an eight-node DGX Spark cluster can support not only large-model serving but also local pre-training of smaller models. NVIDIA said the workflow is intended as an on-ramp: prototype and pre-train locally, then scale out as the model grows. The post did not provide a date for broad availability beyond saying eight-node support in NVIDIA Sync is expected by the end of the year.