NVIDIA Quantum-2 takes supercomputing to new heights and the cloud | NVIDIA press release

Santa Clara, Calif., November 9, 2021 — GTC — NVIDIA today delivers the crushing performance, widespread availability, and robust security demanded by cloud providers and supercomputing centers: Introducing NVIDIA Quantum-2, the -generation InfiniBand network form.

The most advanced end-to-end networking platform to date, NVIDIA Quantum-2 is a 400 Gbps InfiniBand networking platform with NVIDIA Quantum-2 switches, ConnectX-7® network adapters and BlueField-3® data processing unit (DPU). , As well as all the software for this new architecture.

With the introduction of NVIDIA Quantum-2, supercomputing centers will be able to dramatically increase the number of users they accept and make them available to a large number of users outside the organization. At the same time, more and more cloud service providers around the world are beginning to offer supercomputing services to millions of customers.

NVIDIA Quantum-2 implements key features to handle demanding workloads running in both supercomputing centers and cloud supercomputing. Leveraging cloud-native technology, NVIDIA Quantum-2 delivers high throughput performance of 400 Gbps and advanced multi-tenant capabilities to accommodate large numbers of users.

Gilad Shainer, Senior Vice President of Networking at NVIDIA, said, “Today’s supercomputing and public cloud requirements are converging. They deliver the best performance for the HPC, AI, and data analytics challenges of generation, and securely separate workloads. And we have to meet different user traffic requirements. This vision of modern data centers is made possible with the NVIDIA Quantum-2 InfiniBand.”

NVIDIA Quantum-2 performance and cloud-native features
With 400 Gbps throughput, NVIDIA Quantum-2 InfiniBand doubles your network speed and triples the number of network ports. It also triples performance, reduces the number of fabric switches in the data center by a factor of 6, and reduces both power consumption and data center space by 7%.

NVIDIA Quantum-2’s multi-tenant performance isolation ensures that activity from one tenant does not affect other tenants, and congestion control systems that take advantage of advanced telemetry are in high demand among users or workloads. reliable flow.

NVIDIA Quantum-2 SHARP v3™ In-Network Computing technology increases the number of acceleration engines for AI applications 32 times more than previous generations. Advanced InfiniBand fabric management for data centers, including predictive maintenance, is performed through the NVIDIA UFM® Cyber-AI platform.

NVIDIA Quantum-2 has an integrated synchronization system with nanosecond precision that allows synchronization with decentralized applications such as database processing, reducing latency and idle times. Nanosecond synchronization is a new feature that allows cloud data centers to be integrated into communication networks and host software-defined 5G wireless services.

Quantum-2 InfiniBand Switch
At the heart of the Quantum-2 is the new Quantum-2 InfiniBand Switch. With 57 billion transistors on 7 nanometers of silicon, this switch is slightly larger than the NVIDIA A100 GPU with 54 billion transistors.

With 64 400 Gbps ports and 128 200 Gbps ports, various switching systems can implement up to 2,048 ports, more than five times the switching capacity of previous-generation Quantum-1s.

The combination of different network speeds, switching capacities and scalability is ideal for building giant next-generation HPC systems.

NVIDIA Quantum-2 switches are major global players including Atos, DataDirect Networks (DDN), Dell Technologies, Excelero, GIGABYTE, HPE, IBM, Inspur, Lenovo, Penguin Computing, QCT, Supermicro, VAST Data and WekaIO. and system providers.

Quantum-2 ConnectX-7とBlueField-3
The NVIDIA Quantum-2 platform offers two network endpoint options: the NVIDIA ConnectX-7 network card and the NVIDIA BlueField-3 DPU InfiniBand.

With 8 billion transistors in a 7 nanometer design, the ConnectX-7 has twice the data throughput of NVIDIA ConnectX-6, the world’s leading HPC network chip today. In addition, the performance of RDMA, GPUDirect® Storage, GPUDirect RDMA and In-Networking Computing is doubled. ConnectX-7 will be sampled in January.

With 22 billion transistors in a 7 nanometer design, the BlueField-3 InfiniBand features 16 64-bit Arm processors that offload and isolate the data center infrastructure stack. BlueField-3 will be sampled in May.

Leave a Comment