Running Frontier-Class AI Models Across 8 NVIDIA DGX Spark Nodes

NVIDIA Expands DGX Spark to 8-Node Clusters for Local AI Development

NVIDIA has announced official support for eight-node DGX Spark clusters, a significant expansion that scales the platform's unified memory to 1 TB. This move enables AI developers to run and interact with frontier-class models, such as Z.ai's 753B parameter GLM 5.3, directly on local, desk-side hardware. The increased memory and compute capacity are aimed at developers building complex multi-agent workloads and chaining multiple inference calls, tasks that previously required access to larger, centralized compute resources. Support for the new configuration will be managed through NVIDIA Sync's Cluster Assistant and is expected by the end of the year.

Technical Benchmarks Reveal Scaling and Topology Options

The performance of an eight-node cluster is enhanced by a new Spark-optimized collective backend called RoCEnante, which reduces latency for tensor-parallel decode operations on switch-connected systems. Benchmarks show this backend helps deliver a 4.43x decode speedup on eight nodes, compared to 3.05x without it. This allows a model like Qwen3.8-27B to scale from approximately 12 tokens per second on a single node to over 52 tokens per second on an eight-node switched cluster. The choice of network topology also plays a critical role; while a cost-effective four-node ring configuration performs well for single-model inference, a switch is required for eight-node scaling, splitting the cluster into independent groups, or utilizing the RoCEnante backend.

• Total Unified Memory: 1 TB across eight nodes.
• Frontier Model Support: Enables local execution of models like the full GLM 5.3 (753B).
• Inference Scaling: Achieves a 4.43x decode speedup on eight nodes with RoCEnante.
• Training Performance: Reaches 97.4% scaling efficiency for pre-training, with a 7.79x speedup on eight nodes.

Orchestration and Training Broaden System's Impact

Beyond single-model inference, the eight-node configuration is designed for sophisticated multi-model orchestration using tools like NVIDIA NeMo Switchyard. This software allows a cluster to route requests to different models based on task complexity, routing simple queries to a small model and reserving larger models for difficult ones. This approach can increase aggregate throughput by up to 3x while maintaining quality. Furthermore, the platform demonstrates near-linear scaling for training small models from scratch. In a test pre-training a ~900M parameter model, an eight-node cluster achieved a 7.79x speedup at 97.4% efficiency, positioning DGX Spark as a viable system for the entire development cycle, from local pre-training to multi-model deployment.

NVIDIA's expansion of DGX Spark to eight nodes is a strategic move to capture the entire AI development workflow, from local pre-training and experimentation with frontier models to sophisticated, multi-model inference orchestration. By bringing datacenter-class capabilities to a desk-side form factor, the company lowers the barrier to entry for advanced agentic AI development and further entrenches its hardware and software ecosystem.