[C114 Communication Network] Haisiman Jiangwei: Four factors determine the evolution of data center networks
2022-8-22
[C114 Communication Network] Haisiman Jiangwei: Four factors determine the evolution of data center networks
C114 News September 11 (Chen Huajie) Recently, the 22nd China International Optoelectronic Expo kicked off in Shenzhen International Convention and Exhibition Center. Man Jiangwei, customer-side optical module system architect of HiSilicon Optoelectronics Co., Ltd. attended the "New Generation Data Center Optical Interconnection Technology Development and Innovation" forum and said that facing the growth of traffic demand for artificial intelligence, machine learning, public cloud and other services, data communication The network (DCN) needs to consider the evolution of the next-generation large-bandwidth network.
"There are a total of four key elements that determine the evolution of data center networks." Man Jiangwei introduced "Data center users and suppliers promote the development of the standard-driven industry, switching chips determine the evolution of data center network switches, and the evolution of server CPUs and network cards are data The source of central traffic demand, high-bandwidth and low-cost optical interconnection supports the evolution of data centers."
In terms of industry standards, the IEEE 802.3 Ethernet working group established a project in 2013 to formulate the 400G standard, and approved the IEEE 802.3bs 400G Ethernet standard specification in December 2017. At present, the 200GE/400GE standards are mature, and 400GE all scenarios are supported.
In terms of switching chips, the capacity of commercial switching chips has doubled in two years, and larger-capacity switches can be built, requiring a higher-speed optical interconnection solution. The 2*200G/2*400G architecture is the first to obtain the bonus of bit cost, bit power consumption, and highly integrated DC network architecture. Users who are more thirsty for bandwidth will choose one to upgrade to 400G@25.6T and evolve to 400G@51.2T in the future. Some users choose to upgrade to 200G first, and then upgrade to 400G@51.2T in the future.
In terms of servers, the CPU chip rhythm and server shipment forecasts both show a short-lived 50G. After 2020, the momentum of 100G servers has surpassed that of 5G servers. PCle 4.0 network card will be launched in 2020Q3, mainstream I/O will reach 50G, and IO will reach 100G/200G for high-end applications. The PCle 5.0 network card will be launched in 2021H1, and the mainstream I/O will be increased to 100G, and the IO can reach 400G in high-end applications.
In terms of optical interconnection, in order to achieve higher speeds, optical modules generally adopt three approaches: increasing the speed of optical devices, increasing the number of channels, and adopting high-order modulation technology. PAM4 technology introduces short-range modules to effectively increase the speed of optical modules. The speed of 200G is increased by 2 times compared to 100G, and the speed of 400G is increased by 4 times compared to 100G. When the architecture is 4 lanes, the cost per bit and power consumption of 400G modules are lower.
Man Jiangwei said that in the OTT data center network, 100G and 400G are large nodes, and 200G based on a single lane50G is a small node. The 400G solution has better bit cost and power consumption. The 200G solution currently has advantages in absolute cost and power consumption. As the 100G lanes industry matures, 400G FR4 has more cost, power consumption and integration advantages than 200G FR4.
C114 News September 11 (Chen Huajie) Recently, the 22nd China International Optoelectronic Expo kicked off in Shenzhen International Convention and Exhibition Center. Man Jiangwei, customer-side optical module system architect of HiSilicon Optoelectronics Co., Ltd. attended the "New Generation Data Center Optical Interconnection Technology Development and Innovation" forum and said that facing the growth of traffic demand for artificial intelligence, machine learning, public cloud and other services, data communication The network (DCN) needs to consider the evolution of the next-generation large-bandwidth network.
"There are a total of four key elements that determine the evolution of data center networks." Man Jiangwei introduced "Data center users and suppliers promote the development of the standard-driven industry, switching chips determine the evolution of data center network switches, and the evolution of server CPUs and network cards are data The source of central traffic demand, high-bandwidth and low-cost optical interconnection supports the evolution of data centers."
In terms of industry standards, the IEEE 802.3 Ethernet working group established a project in 2013 to formulate the 400G standard, and approved the IEEE 802.3bs 400G Ethernet standard specification in December 2017. At present, the 200GE/400GE standards are mature, and 400GE all scenarios are supported.
In terms of switching chips, the capacity of commercial switching chips has doubled in two years, and larger-capacity switches can be built, requiring a higher-speed optical interconnection solution. The 2*200G/2*400G architecture is the first to obtain the bonus of bit cost, bit power consumption, and highly integrated DC network architecture. Users who are more thirsty for bandwidth will choose one to upgrade to 400G@25.6T and evolve to 400G@51.2T in the future. Some users choose to upgrade to 200G first, and then upgrade to 400G@51.2T in the future.
In terms of servers, the CPU chip rhythm and server shipment forecasts both show a short-lived 50G. After 2020, the momentum of 100G servers has surpassed that of 5G servers. PCle 4.0 network card will be launched in 2020Q3, mainstream I/O will reach 50G, and IO will reach 100G/200G for high-end applications. The PCle 5.0 network card will be launched in 2021H1, and the mainstream I/O will be increased to 100G, and the IO can reach 400G in high-end applications.
In terms of optical interconnection, in order to achieve higher speeds, optical modules generally adopt three approaches: increasing the speed of optical devices, increasing the number of channels, and adopting high-order modulation technology. PAM4 technology introduces short-range modules to effectively increase the speed of optical modules. The speed of 200G is increased by 2 times compared to 100G, and the speed of 400G is increased by 4 times compared to 100G. When the architecture is 4 lanes, the cost per bit and power consumption of 400G modules are lower.
Man Jiangwei said that in the OTT data center network, 100G and 400G are large nodes, and 200G based on a single lane50G is a small node. The 400G solution has better bit cost and power consumption. The 200G solution currently has advantages in absolute cost and power consumption. As the 100G lanes industry matures, 400G FR4 has more cost, power consumption and integration advantages than 200G FR4.
