The full 753B GLM 5.3 model from Z.ai now starts on a desk cluster, and a new collective backend lifts decode speedup.
NVIDIA now supports official eight-node DGX Spark clusters, a configuration that pools unified memory across the machines to reach 1 TB. The larger pool lets developers start models that would not fit on four nodes, including the full 753B GLM 5.3 model from Z.ai, all running locally on desk hardware rather than in a data center.
A single DGX Spark ships with 128 GB of unified memory, four nodes reach 512 GB, and eight nodes reach 1 TB. NVIDIA announced the eight-node setup at GTC Berlin alongside NVIDIA Sync, whose Cluster Assistant handles setup and device management. Sync support for eight-node clusters is coming by the end of the year.
NVIDIA also built a Spark-optimized collective backend called RoCEnante that cuts all-reduce latency for tensor-parallel decode on switch-connected clusters. The backend lifts decode speedup at eight nodes to 4.43x, up from 3.05x without it. Benchmark data shows Qwen3.8-27B in NVFP4 decoding at about 12 tokens per second on one node and 52 tokens per second on eight.
NVIDIA NeMo Switchyard routes requests between model pools on a shared eight-node cluster. Switchyard matched Qwen-only quality on a tool-use benchmark while raising aggregate throughput up to 3x when most traffic shifted to a smaller model. Pre-training a nanochat model from scratch across eight nodes delivered a 7.79x speedup at 97.4% scaling efficiency.
What matters
- NVIDIA announced official support for eight-node DGX Spark clusters at GTC Berlin.
- Builders can place the full 753B GLM 5.3 model on one desk cluster with 1 TB of memory.
- NVIDIA Sync support for eight-node clusters arrives by year end, with Cluster Assistant handling setup.
Why it matters
NVIDIA Sync support for eight-node clusters arrives by year end, with Cluster Assistant handling setup.
This GenAI News article was prepared in original wording using reporting and materials published by NVIDIA Developer Blog. Source reference: https://developer.nvidia.com/blog/running-frontier-class-ai-models-across-8-nvidia-dgx-spark-nodes/.
Drafted by the GenAI News review pipeline.