NVIDIA Networking Buyers Guide

NVIDIA Networking is a sub-brand of NVIDIA that evolved from its purchase of Mellanox Networks in 2020. It is an umbrella term that covers all its switch, NIC and interconnect products.

A network is two or more connected computer devices that communicate and share resources. At large scale such as in traditional datacentres a network usually has three layers - access (connecting to desktops and laptops), aggregation (connecting access switches to the network core) and core (connecting to centralised resources such as servers and storage).

A spine-leaf network is an alternative architecture to the traditional network described in the last question, where the three access / aggregation / core layers are replaced by just two layers - spine / leaf. This simplified structure offers lower latency, better performance and simplified scalability.

Ethernet is a wired networking protocol that connects devices on a local network (LAN) using physical cables. It is used at every scale from home networks to large corporate organisations.

InfiniBand is a alternative wired networking protocol designed specifically for high-performance computing (HPC), datacentres, and AI clusters. It prioritises very high throughput and low latency.

A network switch is a hardware device that connects multiple devices, such as computers, printers, and servers, on a wired local area network (LAN).

A network interface card, or NIC, is a hardware device that is fitted within a PC, workstation, server or storage appliance to connect it to a network switch, in order to share resources such as files, printers, and servers, on a wired local area network (LAN).

A Smart NIC delivers all the connectivity of a standard NIC, but employs remoter direct memory access (RDMA) or RDMA over Converged Ethernet (RoCE) to remove the latency usually introduced by the CPU, system memory and operating system. Learn more by watching our EXPLAINER VIDEO.

A Super NICs is a specialised card optimised to accelerate network traffic for AI workloads. They are built around a high-performance network ASIC, making it a more streamlined and less computationally intensive solution compared to a DPU, so GPU-to-GPU communication is prioritised whilst minimising CPU overhead and latency. Learn more by watching our EXPLAINER VIDEO.

A DPU (data processing unit) is a specialised network card designed to offload, accelerate, and isolate software-defined networking, storage and security functions. This significantly enhances datacentre performance and efficiency and creates a secure, zero-trust environment that streamlines operations. Learn more by watching our EXPLAINER VIDEO.

RDMA, or Remote Direct Memory Access, is a technology that allows data to be transferred directly between the memory of two computers - via an Infiniband NIC - without involving the operating system or CPU on either end.

RoCE (pronounced rocky), or RDMA over Converged Ethernet, is a technology that allows data to be transferred directly between the memory of two computers - via an Ethernet NIC - without involving the operating system or CPU on either end.

A network transceiver is a device that combines a transmitter and a receiver to both send and receive data signals over a network connection. They are essential for network equipment such as switches and routers, converting signals between electrical and optical or copper formats.

A small form-factor pluggable (SFP) module is a type of transceiver that transfers data in a single channel over copper or optical fibre cabling from a network switch to a server or storage device.

A quad small form-factor pluggable (QSFP) module is a type of transceiver that transfers data over four channels over copper or optical fibre cabling from a network switch to a server or storage device.

An octal small form-factor pluggable (OSFP) module is a type of transceiver that transfers data over eight channels over copper or optical fibre cabling from a network switch to a server or storage device.

Multi-fibre push-on (MPO) is a type of connector that removes the need for SFP transceivers. They are used when the photonics element of the fibre-optic are integrated into the ASIC die in very high-end switches.

Power-to-Connector (P2C) and Connector-to-Power (C2P) are airflow standards through the switch chassis. Also know as 'Forward' and 'Reverse' respectively, they are options of how the cooling fans within a switch are configured depending on their mounting in a datacentre rack cabinet.

The management of an NVIDIA network switch is the process of configuring, monitoring, and maintaining the switch to manage data flow, enhance security, and facilitate efficient communication between devices such as GPU-accelerated servers and storage arrays. This can be done via an inbuilt operating system (OS) or through external means or a third-party OS.

Equal Cost Multi-Path (ECMP) is a routing strategy that allows a network device to use multiple paths in local area networks (LANs) with the same cost to forward traffic to a single destination. It enables load balancing, which increases bandwidth utilisation and improves network reliability by distributing traffic across these equal-cost paths.

Multi-Chassis Link Aggregation (MLAG), is a network technology that allows two or more switches to act as a single logical switch for link aggregation, providing redundancy and increased bandwidth. If one switch fails, the other can take over the traffic, minimising downtime and making it ideal for datacentre environments.

Spanning Tree Protocol (STP) is a routing strategy that prevents network loops and broadcast storms in local area networks (LANs) with redundant links. STP works by using an algorithm to build a single, loop-free logical topology, which it achieves by blocking redundant paths while keeping them available as backups. This ensures that there's only one active path between any two devices on the network.

PAM4, or Pulse Amplitude Modulation with 4 levels, is a signalling technique that transmits data using four different signal amplitude levels, allowing for two bits of data to be encoded per symbol, which effectively doubles the data rate compared to traditional two-level (PAM2) or single-level Non-Return-to-Zero (NRZ) systems.