What is NVIDIA Quantum 2?
With features like software-defined networking, In-Network Computing, performance isolation, sophisticated acceleration engines, remote direct memory access (RDMA), and the fastest rates and feeds up to 400Gb/s, NVIDIA Quantum 2 powers the top supercomputing data centers in the world.
NVIDIA Quantum-2 InfiniBand Platform
Outstanding performance at any scale for cloud-native supercomputing.
Record-Breaking Performance in Network Communications
The seventh version of the NVIDIA InfiniBand architecture, NVIDIA Quantum 2, provides the fastest networking speed and feature sets available to scientific researchers and AI developers so they may tackle the most difficult issues in the world. With features like software-defined networking, In-Network Computing, performance isolation, sophisticated acceleration engines, remote direct memory access (RDMA), and the fastest rates and feeds up to 400Gb/s, NVIDIA Quantum-2 powers the top supercomputing data centers in the world.
- Speed of Data: 4X MPI Efficiency
- Enhanced Performance: 3.2X System Capacity of the Switch
- Increased Scalability: 6.5X Better TCO
- Ready for Exascale: 32X More AI Speedup
- Quicker Deep Learning
Enhancing HPC and AI Supercomputers and Applications
Accelerated In-Network Computing
Faster interconnects and more intelligent networks are necessary for today’s high-performance computing (HPC), artificial intelligence (AI), and hyperscale infrastructures in order to analyze data and execute intricate simulations more quickly and effectively. With preconfigured and programmable compute engines, including the third generation of NVIDIA Scalable Hierarchical Aggregation and Reduction Protocol (SHARPv3), Message Passing Interface (MPI) Tag Matching, MPI All-to-All, and programmable engines, NVIDIA Quantum 2 expands and improves its In-Network Computing while offering the best ROI and cost per node.
Performance Isolation
In order to deliver traffic isolations, almost eliminate performance jitter, and ensure predictive performance as though the application were running on a separate machine, the NVIDIA Quantum 2 InfiniBand platform offers creative proactive monitoring and congestion management.
Cloud-Native Supercomputing
The NVIDIA Cloud-Native Supercomputing platform uses high-speed, low-latency NVIDIA Quantum 2 InfiniBand networking in conjunction with the NVIDIA BlueField data processing unit (DPU) architecture. On-demand high performance computing (HPC), data protection, user control and isolation, bare-metal performance, and AI services are all easily and safely provided by the solution.
Delivering Data at the Speed of Light
Adapters
The NVIDIA ConnectX-7 InfiniBand adapter has single or twin network connections at 400Gb/s and comes in a variety of form factors. It supports PCIe Gen4 and Gen5.
Preprocessing data techniques and offloading application control pathways to the network are made possible by the ConnectX-7 InfiniBand adapters’ additional programmable engines and sophisticated In-Network Computing features.
Switches
64 400Gb/s ports or 128 200Gb/s ports on physical 32 octal small form-factor (OSFP) connections make up the NVIDIA Quantum-2 line of switches. Both internal and externally managed air-cooled and liquid-cooled versions of the small 1U switch design are available.
With a capacity of over 66.5 billion packets per second (bpps), the NVIDIA Quantum 2 series of switches provides an aggregated 51.2 terabits per second (Tb/s) of bidirectional throughput.
Transceivers and Cables
The NVIDIA Quantum-2 connectivity choices provide you the most freedom to create the topology of your choice. They consist of various transceivers, active copper cables (ACCs), direct attached cables (DACs) with 1–2 and 1–4 splitter choices, and multi-fiber push-on connectors (MPOs).
Additionally, new 400Gb/s clusters can be connected to existing 200Gb/s or 100Gb/s-based infrastructures via backward compatibility.
World-Leading Networking Performance, Scalability, and Efficiency
Performance
- Each port can have 400 GB of bandwidth.
- A single switch can have 64 400 GB ports or 128 200 GB ports.
- More than 66.5 billion bidirectional packets per second from a single NVIDIA Quantum 2 switch.
Breaking It Own Records
- Compared to the previous generation, twice as much bandwidth per port.
- Three times the switch radix of the prior generation.
- 32X more AI acceleration power per switch and 4X MPI performance compared to the previous generation.
- A four-switch-tier system with more than one million 400Gb/s nodes (three hops).
- DragonFly+ network, which is 6.5 times more advanced than its predecessor.
- Data centre power and space are reduced by 7%.
Important Features
- GPUDirect RDMA, GPUDirect Storage, and full transport offload RDMA.
- Engines for Programmable In-Network Computing (MPI) Hardware acceleration from all to all.
- Acceleration of MPI Tag Matching hardware.
- Advanced adaptive routing, congestion management, and QoS Self-healing networking are features of NVIDIA SHARPv3.
