Nvidia adds eight-machine clustering to DGX Spark
Clusters of 128 GB systems can provide 1 TB of unified memory, but eight-machine configurations currently require manual setup.
Nvidia unveiled expanded clustering for its DGX Spark AI mini PC at GTC Berlin, saying users can link eight machines to provide 1 TB of unified memory when each has 128 GB. Clusters using the 64 GB version can provide up to 512 GB.
The company said the extra memory should allow larger models, extended context windows and several AI agents to run simultaneously. It named GLM 5.3 as a model that could benefit from the 1 TB configuration. Nvidia's announcement did not include benchmarks showing how performance scales across eight systems.
Built-in ConnectX-7 hardware handles communication between the machines. Eight-machine setups must currently be configured manually, but Nvidia plans to support them through Nvidia Sync later in October. A planned Model Launcher feature is intended to simplify compatible model downloads, deployment across connected machines and access from coding tools including OpenCode.
Nvidia also reported gains for specific workloads rather than general performance improvements. It cited up to 30% faster performance on DGX Spark with Qwen3.8 27B and up to 20% higher throughput on DGX Station with GLM 5.3 Flash. Speculative-decoding improvements allow RTX PRO 6000 Blackwell to finish agentic tasks with Qwen 3.8 Flash Next up to 40% faster, according to the company.
New memory-optimized checkpoints include Qwen3.8 Flash Next FP4 and GLM 5.3 Flash FP4. Nvidia says these are designed to lower memory use and preserve capacity for KV cache. The previously announced 64 GB DGX Spark models retain the GB10 Grace Blackwell Superchip and become available October 23.
