NVIDIA adds eight-node DGX Spark clusters
NVIDIA now officially supports eight-node DGX Spark clusters with 1 TB of unified memory, enough to serve Z.ai's full 753B-parameter GLM 5.3 model locally.
NVIDIA announced at GTC Berlin that it now officially supports eight-node DGX Spark clusters. Each node has 128 GB of unified memory, so eight nodes give 1 TB. That is enough for models that would not start on a four-node system, including Z.ai's full GLM 5.3 (753B) checkpoint. NVIDIA Sync's Cluster Assistant handles setup, but eight-node support in Sync is not due until the end of the year.
In NVIDIA's own benchmarks, the full GLM 5.3 decoded at 31 tokens per second for a single user. Total decode throughput rose to 67 TPS at concurrency 8, while prefill held at about 1,600 TPS. GLM 5.3 Flash generated about 80 TPS.
On a dense Qwen3.8-27B (NVFP4) model, decode throughput grew from 11.86 TPS on one node to 52.58 TPS on eight switch-connected nodes, a 4.43x speedup. Prefill scaled only 1.94x. An eight-node physical ring reached 40.11 TPS. NVIDIA says ring setups cost less, but a switch is needed to split a cluster into separate groups.
The switch results use RoCEnante, an all-reduce backend that is now merged into FlashInfer. Instead of passing data around a ring, every node sends its data directly to every other node. This cuts latency for small messages but moves more data overall. NVIDIA reports about 21% higher decode throughput at four nodes and 18% at eight. The backend works only on switch-connected setups and falls back to NCCL, which stays the default for larger clusters and higher concurrency. It builds on a one-shot all-reduce approach contributed by Jason Cook (el8).