Two DGX Stations split one model across distinct GPU memories
NVIDIA’s deployment guide serves a roughly 464 GB GLM-5.3 checkpoint across two DGX Stations because it exceeds one GB300 GPU’s 252 GB memory. Tensor parallelism splits weights and computation across the GPUs, connected through RDMA, while one station exposes the API. The memories remain distinct: distributing a model adds communication requirements rather than creating a single larger GPU.
NVIDIA Technical Blog · Nov 13, 2026
Why it matters
A model’s active parameters per token do not determine how much storage all its weights require.
More context
The documented setup needs two Linux GB300 Stations, validated RDMA connectivity and at least 464 GB checkpoint storage on each machine. Review licenses and the pinned vLLM configuration, then benchmark representative prompts rather than assuming peak aggregate throughput predicts individual latency.
Keep in mind
A vendor tutorial for costly, specific hardware, not a general multi-GPU compatibility guarantee. Sparse expert activation saves computation but does not remove the need to make the full weights available.
Original read