What is a rack-scale system?
A rack of independent servers is not the same as an integrated rack-scale system. In the latter, compute, connections, power, and cooling are designed together.
One documented NVL72 arrangement
Each green square represents one GPU.
One real example
NVIDIA’s documented NVL72 arrangement has 18 compute trays, each with four GPUs and two CPUs. Nine NVLink switch trays connect the GPU fabric.
That is 72 GPUs and 36 CPUs across the rack, not inside one ordinary server.
Why the boundary matters
A closely connected domain can help communication-heavy jobs. Software must still divide the work across devices.
Shared infrastructure can also affect several trays at once. Some failures isolate a component; others require coordinated recovery. Check the particular system’s documented behavior.
Do not assume a rack-scale system has one operating system, one memory allocation, or one universal restart procedure.
Try it yourself
Calculate GPUs per rack, then identify three dependencies that could affect several trays.
Check your answer
18 × 4 = 72 GPUs. Fabric, power, and cooling can be shared dependencies. Their actual failure impact depends on the installed redundancy and protection design.