Nvidia gave its Infiniband networking fabric a speed boost with the launch of the Quantum-2 switch silicon at GTC 2021.

“The network is a vital central nervous system of the computer,” CEO Jensen Huang said during his keynote. “The network connects thousands of GPUs into a giant supercomputer, determines the scalability, and ultimately the performance.”

The Quantum-2 Infiniband switch silicon — based on technology acquired with the purchase of Mellanox in 2019 — effectively doubles the throughput over the previous generation and offers 64, 400 Gb/s ports or up to 128, 200 Gb/s ports.

Enabling this is a massive processor die, which packs 58 billion transistors into a package slightly larger than Nvidia’s A100 GPU.

In addition to boosting the throughput and switching capacity, the new chip is Nvidia’s first with support for multi-tenant performance isolation, which enables operators to network multiple workloads simultaneously across the fabric.

“Quantum-2 is the first networking platform to offer the performance of a supercomputer and the share-ability of cloud computing,” Huang said. “Before Quantum-2, you either got bare-metal high performance or multi-tenacy — never both.”

The technology employs a cloud-native, telemetry-based congestion control system to prevent activity from one tenant from disrupting others.

The switch also features a new nano-second timing system that synchronizes distributed applications, like database processing, to reduce wait and idle overheads.

“Nanosecond timing will also allow cloud data centers to become part of the telecommunications network and host software-defined 5G radio services,” Huang said. “This networking speed, switching capacity, and scalability is coming just in time for the giant HPC systems that the world needs to build.”

Quantum-2 is available now from Nvidia’s partners, including Dell Technologies, IBM, Inspur, Lenovo, and Hewlett Packard Enterprise, to name a handful, and can power switch systems with up to 2,048 interfaces at 400 Gb/s.

Quantum-2 is the latest complement to Nvidia’s growing networking ecosystem. The switch is designed to interface with servers equipped with the company’s upcoming ConnectX-7 NIC cards and BlueField-3 data processing units (DPUs).

Nvidia announced its third-generation BlueField DPU and at the company’s spring GTC conference in April.

BlueField-3 is designed to offloads data center services — networking, storage, and security functionality — at line rates up to 400 Gb/s, freeing the CPU to run business-critical applications and alleviate bottlenecks.

Alongside BlueField-3, Nvidia also announced ConnectX-7, which features up to four, 400 Gb/s interfaces in a PCIe Gen 5.0 NIC form factor. “CX-7 is the fast NIC ever built,” Huang boasted.

ConnectX-7 and BlueField-3 will begin sampling in January and May respectively.