Set microgrid news, nowadays, DPU is becoming an important promoter of the era of multi-chip integration, since the CPU and DPU joined the chip matrix, NVIDIA proposed a new computing architecture for 3U integrated data centers, which breaks the original von Neumann architecture-based traditional data center, so that computing around the data according to local conditions.
At the same time, network technology also needs to make new changes at the same time. How to make the network intelligent has become one of the key technologies of today's data-centric architecture.
"A lot of data transfer, pre-processing, can bypass the CPU to make data more efficient to go where it should go, making the network intelligent is the change that is currently happening on the network side." Meng Qing, marketing director of NVIDIA Network, recently pointed out this to the media, including Jiwei Network, at a media exchange meeting.
These resulting changes include the birth of DPU, In-Network Computing, and the collaboration of DPUs and GPUs to use AI to ensure that network traffic is not attacked (Morpheus solution).
Networking is happening
Jiwei Network understands that NVIDIA's new computing architecture of data centers can reduce communication delay by more than 10 times, and behind this, network computing has become a key supporting technology. NVIDIA stressed that the network is a bridge that can be built, and modern data center infrastructure needs modern networks.
According to Synergy, there are currently 314 new hyperscale data centers known, and the operating data center installation base will exceed the 1,000 mark within three years and continue to grow rapidly thereafter. With the rapid growth of data centers, the global network hardware market has also maintained rapid growth. According to IDC data, in 2021, global Ethernet switch revenue increased by 11.8% year-on-year, router market revenue increased by 7% year-on-year, Ethernet switch market increased by 9.7%, and router market increased by 6.5%.
Behind the rapid growth is the continuous landing and development of 5G, AI, cloud computing, data centers and other applications, pushing up the demand for network technology and equipment.
"Networks will have different characteristics depending on the function of the data center, and the common characteristics are high bandwidth, low latency, data security, and performance isolation – this is the direct demand for the network in the modern data center." Meng Qing pointed out that the current data center is developing in the direction of infrastructure platforms, which are used to support applications from people's livelihood to scientific exploration, artificial intelligence, and private services. The data center itself is also evolving from a comprehensive data processing computing center to many branches, such as a cloud center that supports mobile Internet, games, APP, etc., AI training for speech analysis, image recognition, etc., and also for big data storage. These data centers have different divisions of labor in various Internet factories, and the "east and west counting" proposed by the state is a typical policy adjustment on cost and energy for this trend.
Meng Qing, Marketing Director of NVIDIA Network
Jiwei Network noted that the current NVIDIA network acceleration platform has two solutions, InfiniBand and Ethernet, of which there are DPUs. NVIDIA's definition of the role of DPUs in the data center is to accelerate data-intensive tasks, including networking, security and storage, thereby enabling a software-defined, hardware-accelerated data center infrastructure.
It can be seen that the hardware role of DPU in NVIDIA network technology can be seen, and its addition is the core support for the realization of modern networks.
For what work NVIDIA will do in terms of DPU this year, Meng Qing said that first of all, NVIDIA will do its best to do relevant work around the current BlueField-2 DPU to help customers realize the benefits of DPU and can actually deploy to the production environment. Second, NVIDIA will continue to focus on leveraging DPUs, the DOCA software environment and related talent.
"The demand for computing power in AI and data centers is unlimited, and the same is true for network speeds, just as we expect faster computers, and the efficiency of infrastructure can lead to a lot of software and application innovation." Meng Qing said.
Reinventing hardware makes networks smarter
Cui Yan, a network expert at NVIDIA, pointed out at the media briefing that data centers are growing exponentially, and the requirements for network bandwidth at the application and server levels are getting higher and higher, while providing better security and strong functionality. To meet these requirements, an Ethernet platform that can provide high performance, low latency, and advanced virtualization and simulation support in large-scale application scenarios is an essential solution.
Nvidia network expert Cui Yan
Nvidia unveiled the next-generation InfiniBand networking platform Quantum-2 and the next-generation Ethernet platform Spectrum-4 at gtclic conferences last year and this year, respectively.
Spectrum-4 is the world's first 400Gbps end-to-end network platform with 51.2Tbps switching throughput up to 4x higher than previous generations. The platform consists of the NVIDIA Spectrum-4 switch series, ConnectX-7 smart NICs, NVIDIA BlueField-3 DPU, and DOCA data center infrastructure software to dramatically accelerate large-scale, cloud-native applications.
Cui Yan introduced that Spectrum-4's main application scenarios are in large-scale cloud computing, enterprise AI and simulation, which can provide better network performance optimization for these scenarios.
He stressed that Spectrum-4 is not only a network platform, it can be integrated with other NVIDIA platform software and applications, providing better Ethernet connectivity and support for upper-layer applications at the bottom layer. In the modern data center, workloads can be optimally targeted, secure, low-latency and adaptive routing, and complex congestion control mechanisms.
The Quantum-2 platform consists of a Quantum-2 switch, a ConnectX-7 network card, a BlueField-3 DPU, and all the software that supports this new architecture. The platform doubles network speeds and triples performance while reducing data center energy consumption and space by 7% each.
In response to the similarities and differences between the two acceleration network platforms and the applicable scenarios, Cui Yan pointed out in an interview with the media including the microgrid that both the Spectrum-4 and Quantum-2 platforms are an important part of the network infrastructure. The difference is that Spectrum-4 is more suitable for cloud computing data centers, enterprise AI and simulation, and Quantum-2 is more suitable for high-performance computing, artificial intelligence, especially large-scale GPU clusters, as well as various biomedical analysis, geological climate prediction analysis and other large-scale data analysis.
He adds that for cloud-native applications, the boundaries between the two network structures will gradually blur. Taking the digital twin as an example, the network and the computer use precision timing protocols for synchronization, but also need to reduce the delay of packet transmission to a greater extent, the network is the core backbone.
"Data centers are being transformed into AI factories that process massive amounts of data to achieve intelligence," Jen-Hsun Huang, CEO and founder of NVIDIA, put forward his own thinking at GTC 2022. The era of multi-chip integration carries the computing power required for high-performance computing, AI, etc., we can easily focus on innovative hardware, but it is also easy to ignore some hidden existence, how to minimize the delay, how to deal with the exchange of large-scale data, and how to make data transmission more efficient and more intelligent Is the need for network technology to continue to change, thus supporting the transformation of data centers and becoming a solid bridge.
(Proofreading/Sharon)