Skip to content
GIRIVON
Infrastructure+
Systems+
Components+
Computing Devices+
Solutions+
Compute Tools+
Knowledge+

GPU Clusters

Interconnected multi-GPU servers designed to run as a single resource for parallel training, simulation and rendering.

Scale-up and scale-out

Inside a server, GPUs communicate over NVLink/NVSwitch or Infinity Fabric (scale-up) or over PCIe. Between servers, traffic crosses the network fabric (scale-out) — InfiniBand or RDMA-capable Ethernet. Collective operations such as all-reduce are bounded by the slowest link in that path, so the fabric is sized from the gradient or activation traffic of the workload.

GPU memory does not simply add up

Eight 80 GB GPUs do not behave as one 640 GB device. A model larger than one GPU's memory has to be split using tensor, pipeline or fully-sharded data parallelism, and each method adds communication and per-GPU overhead. The GPU VRAM calculator shows per-GPU memory under these parallelism settings.

Facility checks

Dense GPU servers commonly draw multiple kilowatts each. Check rack power per phase, PDU capacity, cooling capacity per rack and floor loading before committing to a cluster layout — the rack power and cooling calculators do this arithmetic.

Specifications to compare

  • GPU count
  • Memory per GPU
  • Total GPU memory
  • Interconnect
  • Network speed
  • Max system power

Calculate before you buy

Typical workloads