Enable cloud-based collaboration on AWS to provide quantum computing resourcesBraket ServicesLater, at the European GTC 2021 event, NVIDIA announced the NVIDIA Quantum-2, a network platform built with a new generation of InfiniBand switches. It will be able to use ConnectX-7 NIC or BlueField-3 DPU InfiniBand design, combined with software that supports the new architecture, to correspond to 400Gbps of data transmission capacity, allowing cloud service providers in various regions of the world to provide supercomputing-level cloud collaboration services through this network platform.
Currently, more and more supercomputing resources are being made available to private organizations. However, due to limitations such as location and availability, these computations often rely on remote network connections to transmit data and return analysis results. NVIDIA's proposed Quantum-2 network platform aims to achieve the operational fluidity of dedicated servers (bare metal) by supporting data transmission at a scale of 400Gbps, while also ensuring data transmission security and adapting to cloud-native application models.
The Quantum-2 network platform is designed with InfiniBand switches, ConnectX-7 NICs, or BlueField-3 DPU InfiniBand, and is built with software that supports the new architecture. It can meet the stringent workload requirements of supercomputing operations and can even accommodate multi-user access usage models simultaneously.
At the same time, the significant increase in data transmission volume not only doubles the transmission speed, but also triples the overall transmission efficiency. This significantly reduces the number of switches originally required in the data center, reduces the space occupied by switches in the data center by 7%, and reduces power loss by 7%. The number of external network connection ports can be tripled, which means that the Quantum-2 network platform can significantly improve the operating efficiency of the data center, the number of external connections, and significantly reduce operating power consumption.
NVIDIA also emphasized that SHARP v3 In-Network Computing technology will enable AI computing to be accelerated by 32 times compared to previous products. The UFM Cyber-AI platform will also enable data centers to have InfiniBand fabric management capabilities and add predictive maintenance functions. Multi-user performance isolation design ensures that a single user's computing project will not affect the usage rights of other users. Congestion control telemetry systems and cloud-native functions also ensure reliable transmission, thereby balancing the operation of different computing workloads.
The nanosecond-precision timing system integrated into the Quantum-2 network platform can operate synchronously with distributed applications such as database processing, thereby reducing waiting or idle time. It can also be used in software-defined 5G radio services hosted by telecommunications services.
The InfiniBand switch used in the Quantum-2 network platform is designed using TSMC's 7nm process and is equipped with a total of 570 billion transistors. It is slightly larger than the NVIDIA A100 GPU and is equipped with 64 ports corresponding to 400Gbps data transmission rate, or 128 ports corresponding to 200Gbps data transmission rate, for a maximum of 2048 ports, which is about five times more than the previously launched Quantum-1.
Companies including Atos, DataDirect Networks (DDN), Dell, Excelero, Gigabyte, HPE, IBM, Inspur, Lenovo, Penguin Computing, QCT, Supermicro, VAST Data and WekaIO have already launched server products using the Quantum-2 network platform.
For network endpoints, the ConnectX-7 NIC or BlueField-3 DPU InfiniBand can be used. The former will offer twice the performance of the ConnectX-6 and is expected to begin sampling in January, while the latter will follow in May. NVIDIA also announced the release of the DOCA 1 transport protocol, which incorporates zero-trust protection technology developed in collaboration with software vendors such as Check Point, Juniper, Fortinet, Palo Alto, Trend Micro, Guardicore, VMware, and F5 to ensure data transmission security.
