16x DGX Spark Cluster Hits Line Rate

This is a memory-first AI rack, not a conventional GPU cluster. The interesting part is how far NVIDIA’s desktop-class boxes can be pushed when you treat them as modular building blocks for prefill-heavy workloads.

• –DGX Spark’s 128GB unified memory and 200GbE networking make it a surprisingly coherent node for large-model inference and orchestration
• –The setup work is nontrivial: shared users, SSH, jumbo frames, addressing, and automation matter as much as the hardware once you scale to 16 nodes
• –The proposed prefill/decode split is sensible; it maps heavyweight parallel work to the Sparks and leaves decode to denser boxes later
• –Line-rate networking is a good sign, but software partitioning, scheduler design, and KV-cache placement will decide whether this is elegant or just expensive
• –Compared with H100 or GB300, the value proposition is ecosystem consistency and aggregate memory capacity, not absolute throughput per dollar