让1.6T以太网的思想变为行动

Turning Ideas into Actionfor 1.6T Ethernet Turning Ideas into Action for 1.6T Ethernet Ushering in the Next Ethernet Rate The explosion of Edge computing, virtual reality (VR)and augmented reality (AR), and AI applicationsare set to significantly increase demand from theirnetworks as new requirements accelerate theadoption of higher speed Ethernet connectivitysolutions to satisfy these data needs. This trend isdisrupting traditional timelines for the developmentand availability of connectivity technologies.This will have a major impact on the ecosystempowering these new technologies, including NetworkEquipment Manufacturers, Data Center operators,interconnect manufacturers and system integrators. In our latest “AI Networks for AI Workloads”Advanced Research Report, we predict that by2025, most ports in AI networks will operate at800 Gbps, with a swift shift to 1600 Gbps by 2027.This demonstrates a rapid uptake of the highestspeeds available in the market. This pace ofmigration is nearly twice as fast as what we havetypically observed in traditional front-endnetworks connecting general-purpose servers. Sameh Boujelbene,Vice President at Dell’Oro Group Key Requirements of Next-Gen Applications For AI/ML training and inference, there is a highdegree of sensitivity to bandwidth and latency. Thejob completion times (JCT) are gated by tail latencywhich negatively affects the ROI on the expensiveAI infrastructure because the GPUs will now haveto sit idle waiting for the delayed packets to arrivein order to process them while other jobs in queuesuffer. The IEEE’s 802.3dj group is formulating theupcoming Ethernet standard, which includesphysical layers and management parameters forspeeds up to 1.6 Terabits per second. Since the AItraffic workloads are characterized by the presenceof many elephant flows, the greater the availablebandwidth, the better the large-scale datatransfers can take place which will help reduce jobcompletion times. In addition, the cost per bit, which is a majorconsideration for large deployments, will decreasedue to more efficient optical module designsexpected for 1.6T, resulting in a decreased powerconsumption per bit. Technology Challenges on the Path to 1.6T Ethernet losses and reach useful distances. The channelbetween the ASIC and front panel can be improved byusing new PCB materials, adding more retimers, usinglow-loss flyover cables, or even using co-packagedoptics. In addition, new forward error correction andlink training algorithms are under consideration toimprove overall link fidelity budget. The channelsbetween Ethernet ports will similarly need to adoptnew technologies and techniques to offset the ever-increasing power consumption inherent in fastersymbol rates. As each new Ethernet generation increased the linkspeed, there was a corresponding increase in therate of the signals carrying the data. In some cases,industry has been able to keep the line rate the samebetween successive generations to a certain degreeby utilizing methods such as parallel lanes and PAM4modulation, but continued progress inexorably calls forfaster and faster bit transitions. IEEE 802.3dj gets to 1.6T Ethernet by doubling thephysical layer interface bit rate from 800G technology;112 Gbps up to 224 Gbps. Since PAM4 will continue tobe used, the doubling of bit rate inherently doubles thespectral content of the physical layer signal. Doublingthe symbol rate of the signal requires a combinationof improved analog channel characteristics andmore sophisticated SERDES transmitter and receivertechnologies. As with many communication system advances,carrying 1.6T information across a physical mediumturns into an engineering challenge of how tobest achieve the performance requirementswhile mitigating complexity. Dollar costs, powerconsumption, and latency all need to be factored intothe complexity equation for 1.6T. To achieve the rightbalance for multiple applications, a wide variety ofsolutions will likely be required. Electrical channels like PCB traces, connectors, andcopper cables will be challenged to overcome insertion The Evolving Connectivity Landscape training is typically used to optimize the transmittersand receivers at both end points. However, withincreased data rates to 112 Gbps per lane and nowup to 224 Gbps per lane for 1.6T, the electrical signallosses limit how far even advanced SERDES can reach.Losses can be reduced by using thicker copper wire,but this approach has its drawbacks. Thicker cablesare heavier, harder to manage, and tend to block offthe airflow to the faceplate, which can be problematicfor high density deployments. As a result, for 1.6T,the longest DACs may no longer be practical or evencapable of reaching from the top of rack to the bottom. Connections between Ethernet ports have historicallybeen made through copper cables for shorterdistances and optics for longer reaches. However,starting with 800G the growing disparity in reach, cost,and power between DACs and