NVIDIA Expands DGX Spark Clusters to Eight Systems With One Terabyte of Unified Memory

NVIDIA has expanded DGX Spark clusters to eight systems, pairing up to one terabyte of unified memory for developers running larger AI models on local hardware.

Each DGX Spark used in the eight-system configuration has 128GB of unified memory. Four systems combine 512GB, while eight provide one terabyte. The expanded setup builds on the earlier two-system DGX Spark configuration, which combined 128GB for models of up to 200 billion parameters.

Clusters can run Z.ai's complete GLM 5.3 model, which has 753 billion parameters. Its NVFP4 checkpoint is about 470GB. NVIDIA's test of GLM 5.3 Flash across eight systems reached about 80 generated tokens per second.

In NVIDIA's Qwen3.8-27B NVFP4 benchmark at concurrency one, an eight-system switch setup recorded 5,105 prefill tokens per second and 52.58 decode tokens per second. A single system reached 2,635 prefill tokens per second and 11.86 decode tokens per second. The test used an 8,192-token input and a 256-token output; an eight-system physical ring delivered 40.11 decode tokens per second.

Network layout affected the results. The RoCEnante configuration works with switch-connected systems. NCCL remains the supported default, and NVIDIA recommends it as cluster size and concurrency increase.

A separate NVIDIA training test used a nanochat model with about 896.5 million parameters and a global batch of 524,288 tokens per step. Throughput reached 68,079 tokens per second on eight systems, compared with 8,741 on one system. NVIDIA calculated the eight-system result as 7.79 times the single-system throughput, with 97.4% scaling efficiency.

Eight-system clusters can be configured manually now. NVIDIA plans to add support for them in its Sync software.