Skip to content

NVSwitch

πŸ“ Short definition

NVIDIA chip that links all GPUs in an HGX node via NVLink, providing uniform total bandwidth between every GPU pair.

πŸ“– Detailed definition

In a typical multi-GPU workstation on PCIe, GPUs communicate through the CPU bus β€” slow and asymmetric. NVSwitch replaces that topology with a non-blocking fabric: each GPU can send and receive data at full NVLink speed simultaneously to all others.

An HGX B200 node (8 GPUs) embeds several NVSwitch chips that give each GPU 1,800 GB/s of bidirectional NVLink bandwidth to the others, i.e. ~14.4 TB/s aggregate across the 8 chips β€” a prerequisite for efficient Tensor Parallelism.

NVLink vs NVSwitch: NVLink is the protocol/cable; NVSwitch is the routing chip that forms the complete fabric.

πŸ’‘ Why it matters for on-prem AI

NVSwitch is the architectural boundary between a desktop multi-GPU workstation (PCIe, slow communication) and a true datacenter node (HGX, uniform NVLink fabric). It is what makes Scenario D possible.

⚠️ Common pitfalls

  • Two GPUs linked by NVLink without NVSwitch (point-to-point topology) do not have the same bandwidth as in a full NVSwitch fabric.
  • Absent from all consumer cards β€” reserved for professional HGX/DGX lines.

πŸ”— See also

Give my opinion

Everything is optional: one answer is enough. This site sends nothing: you send it, the way you choose.

Your noteNo note
Your opinion is about
About what? (several choices possible)
What will be sent
Send by e-mail

Private, no account: opens your mail program with the message ready, for ia-on-prem@damien.becherini.fr.

Open a GitHub issue β†—

Public, with a GitHub account. You read the issue again before you publish it.

To paste it wherever you want.

Give at least a note, a subject or a message to be able to send.