Eight DGX Spark nodes become one desk-side cluster: 1 TB of unified memory puts a 753B GLM 5.3 checkpoint on local hardware
Synopsis
NVIDIA announced at GTC Berlin official support for eight-node DGX Spark clusters, raising unified memory from 128 GB on a single node to 1 TB and letting frontier-class models that would not start on a four-node system load locally. The full GLM 5.3 (753B) checkpoint runs on the eight-node cluster, GLM 5.3 Flash generates at roughly 80 tokens per second, and decode throughput on dense Qwen3.8-27B (NVFP4) rises from about 12 TPS on one node to 52 TPS on eight.
Interpretation
Desk-side AI development hardware can now link eight units into a local cluster with 1 TB of unified memory, enough to load and interactively serve frontier-class models. A single DGX Spark provides 128 GB of unified memory and four nodes reach 512 GB; eight nodes reach 1 TB, opening the door to models that previously would not start on a four-node system. The full GLM 5.3 (753B) checkpoint runs across the 8x 128 GB cluster, and GLM 5.3 Flash measured roughly 80 tokens per second in generation on an eight-node cluster.
Decode throughput scales strongly with node count while prefill grows more slowly, and the best topology depends on the workload. On dense Qwen3.8-27B (NVFP4) without speculative decoding or MTP, one node decodes at about 11.86 TPS and eight nodes on a switch reach 52.58 TPS. In the same benchmark, eight nodes in a physical ring reach 40.11 TPS decode versus 52.58 TPS on the switch (RoCEnante) topology, a 4.43x decode scaling; prefill rises from 2,635 TPS on one node to 5,105 TPS on eight switch-connected nodes.
A Spark-optimized collective backend cuts communication latency in tensor-parallel decode and is merged into FlashInfer. It replaces ring all-reduce's multi-round forwarding with one all-to-all exchange plus local reduction, reducing communication steps at the cost of each rank sending its full input to every peer. At four nodes (NVFP4) it gives roughly 5% higher prefill and 21% higher decode throughput, and about 18% higher decode throughput at eight nodes on the non-quantized model; it falls back transparently to NCCL if its startup test fails.
Routing requests by difficulty to models of different sizes raises aggregate throughput on a shared cluster while holding quality. The prototype uses OpenBMB's MiniCPM5 2B for simple tasks and Qwen3.8 Flash Next for complex ones, with a lightweight classifier routing each request or session and a load balancer spreading traffic across replicas. On the 92-case tool-use benchmark including hard cases, the routed agent scored 91/100, matching the Qwen-only configuration; across eight nodes with three Qwen replicas and two MiniCPM replicas, routing half the requests to Qwen gives about 1.5x aggregate throughput and routing a quarter gives about 3.0x.
Perspective
The demonstrated scope is clusters of DGX Spark nodes: inference benchmarks use dense Qwen3.8-27B (NVFP4) and the GLM 5.3 family, and the pre-training experiment trains an approximately 896.5M-parameter nanochat model from scratch. The Spark-optimized collective backend applies only to switch-connected topologies with direct point-to-point connectivity between every node, and NCCL remains the supported default.
The material states that the backend's extra traffic becomes more expensive as messages or node counts grow, and that NCCL performs better for high concurrency, larger clusters or ring topologies; eight-node cluster support in NVIDIA Sync is coming by the end of the year. Near-linear pre-training scaling was measured with each node holding a full model copy and only gradient synchronization once per step.