Nvidia announces new InfiniBand networking hardware

Networking equipment was the news of the day at Nvidia’s GPU Technology Conference (GTC), with new hardware for improved end-to-end performance.

Nvidia announced the Quantum-2 platform, a 400Gbps InfiniBand networking platform consisting of the Quantum-2 switch, ConnectX-7 network adapter, BlueField-3 data processing unit (DPU), and all the software to support the new architecture.

At 400Gbps, NVIDIA Quantum-2 InfiniBand doubles the network speed and triples the number of network ports over the Quantum-1 product. With a three-fold performance increase in performance, data-center fabric switches can be reduced by six-fold, cutting data center power consumption and reducing the overall data center space by 7%, the company says.

“The network is the vital central nervous system of the computer,” said CEO Jen-Hsun Huang in his keynote speech. “Quantum-2 is the first networking platform to offer the performance of a supercomputer and the shareability of cloud computing. This has never been possible before.”

Before Quantum-2, customers could choose either bare metal high-performance or secure multi-tenancy, but not both. “With Quantum-2, your valuable supercomputer will be cloud native, and far better utilized. Performance isolation keeps the activity of one tenant from disturbing others,” he said.

NVIDIA Quantum-2’s multi-tenant performance isolation separates the activity of one tenant from disturbing others, known as “noisy neighbor syndrome.” It does this through a telemetry-based congestion-control system to ensure reliable throughput even when there are spikes in users or workload demands.

The Quantum-2 platform is built around a new InfiniBand switch, which has 57 billion transistors on seven-nanometer silicon. With more than five times the switching capacity of Quantum-1, the new Quantum-2 platform features 64 ports at 400Gbps or 128 ports at 200Gbps. Quantum-2 will be available in configurations of up to 2,048 ports. Thanks to Quantum-2 SHARPv3 In-Network Computing technology, AI-application performance will improve 32 times.

Quantum-2 is available from infrastructure and system vendors including Atos, DataDirect Networks (DDN), Dell Technologies, Excelero, GIGABYTE, HPE, IBM, Inspur, Lenovo, Penguin Computing, QCT, Supermicro, VAST Data, and WekaIO.

New Mellanox hardware

In addition to the Quantum-2 switch, Nvidia introduced new endpoint hardware. ConnectX-7, the latest in its smart host channel adapter (HCA) card, which Huang said doubles the data rate of the ConnectX-6. With its 256-thread data-path processor it does crypto at line rate.

In tandem with the Quantum-2 architecture, ConnectX-7 offers 400Gbps of throughput and Nvidia’s In-Network Computing acceleration engines to provide additional acceleration and scalability for supercomputing, artificial intelligence, and hyperscale cloud data centers.

The second piece of the puzzle is the BlueField-3 data processing unit (DPU) that goes on the ConnectX card, with 400Gbps of Ethernet or NDR 400Gbps InfiniBand network connectivity. BlueField-3 DPU offloads, accelerates, and isolates software-defined networking, storage, security, and management functions with 16 64-bit ARM CPUs.

Those infastructure chores consume 30% of CPU usage, and that number is growing, Huang said. “For multi-billion dollar data centers, the freed up capacity can be a giant cost savings or throughput boost,” he said.

Huang announced major firms plan to provision their next generation firewall services on BlueField: Checkpoint, F5, Fortinet, Juniper Networks, Guardicore, Palo Alto Networks, Trend Micro and VMware. “There are now 1,400 developers working with BlueField and now, cybersecurity companies on Bluefield can provide Zero Trust security as a service,” he said.