4-GPU dedicated node
720 GB total GPU memory with 8% off every card. One tenant, one fixed 7- or 30-day period, one predictable total.
Aggregate resources
Resources scale with each dedicated card. Networking and topology must be verified before a real reservation.
4× 180 GB gives 720 GB across the cards. The application must explicitly distribute its model or work. Independent replicas, tensor parallelism and sharded training have different communication requirements.
Before treating this catalog configuration as a connected machine, confirm visible GPUs, PCIe/NUMA placement, interconnect, driver and collective-library versions. Card count alone does not establish an NVLink or NVSwitch fabric.