NVIDIA DGX Spark Expands Memory for AI Workloads

Until now, NVIDIA DGX Spark has supported scaling up to two nodes, increasing the available memory from 128 GB on one node to 256 GB on two nodes. This capability has now been increased to up to four DGX Spark nodes. DGX Spark also now supports several execution topologies, each tailored to different goals through the low latency of RoCE communication enabled by ConnectX-7 NICs.

- One DGX Spark node: Ideal for low latency, large context size inference, fine-tuning up to 120B parameters, and local agentic workloads - Two DGX Spark nodes: Balanced scaling for faster fine-tuning and larger models, as well as support for up to 400B-parameter inference - Three DGX Spark nodes in a ring: Ideal for fine-tuning larger models or small training jobs - Four DGX Spark nodes with RoCE 200 GbE switch: Local inference server ideal for state-of-the-art models up to 700B parameters, communication intensive workloads, and local AI factory operations Inference can scale up linearly on DGX Spark when internode communication is minimal. When work is largely independent per GPU, the results are aggregated once at the end rather than continuously. In this case, DGX Spark nodes can run in parallel with low synchronization overhead.

For example, a reinforcement learning (RL) workload in NVIDIA Isaac Lab can run many simulations independently on each node. Results are collected in a single step, yielding near-linear scaling across multiple DGX Spark nodes. Inference scaling is less than linear when the workload requires frequent, fine-grained communication between nodes.

During LLM inference, model execution occurs layer by layer, with continuous synchronization required across nodes. Partial results from different DGX Spark nodes must be exchanged and merged repeatedly, which introduces significant communication overhead.