Resultados 1 a 2 de 2
  1. #1
    WHT-BR Top Member
    Data de Ingresso
    Dec 2010

    [EN] Data Center Calls for 800GE Spec

    Arista founder forecasts vendors will ship five million 100GE Ethernet ports this year and double those shipments in 2018. Hard on its heels, 400G shipments could rise to one million units in 2019 and four million the following year.

    Rick Merritt

    Bandwidth-hungry data centers need a fast-track effort this year to define 800 Gbit/second Ethernet links, said a networking veteran. The existing IEEE process is too slow to serve the needs of Web giants, said Andreas Bechtolshiem, chairman of Arista Networks and a serial entrepreneur.

    Network bandwidth has long been the bottleneck for companies such as Amazon, Facebook and Google, trying to connect thousands of servers to handle a flood of Web and mobile traffic. They are moving to 100GE connections this year and will start buying in volume as early as 2019 the 400G systems that are still in the lab today, he predicted.

    At this pace, data centers won’t be able to wait for a formal IEEE process that could take three years to define 800GE, said Bechtolsheim. He called for serdes makers to forge a multi-source agreement this year on 112G interfaces using PAM-4 modulation as the basis for doubling the data rate of the IEEE 400G standard about to be formally ratified.

    “We think the cloud industry needs faster networks and this is the best way to get there,” he said in a keynote at the Linley Cloud Hardware Conference here.

    Bechtolshiem forecasts vendors will ship five million 100GE Ethernet ports this year and double those shipments in 2018. Hard on its heels, 400G shipments could rise to one million units in 2019 and four million the following year, he estimated.

    An unexpected drop in 100G prices triggered the fast ramp, now generating shortages in optics given a fragmented market for different form factors of modules and types of fiber optic cables.

    Analysts “assumed the wrong cost model” and predicted a “100G trickle” this year. “People assumed they had time to ramp…[but] the industry has never seen such a ramp with optics,” said Bechtolsheim, showing revised forecasts as well as his own estimates.

    “The standards have fallen behind demand from the hyper-scalers,” said Bob Wheeler, a principal networking analyst at the Linley Group. “800G would be the next logical step, and I think people are not aware how fast it’s coming,” he said.

    Ethernet chip and system vendors heard calls for Terabit Ethernet from Facebook and Google as early as 2011. They claimed both technical and economic hurdles prevented them from pressing the roadmap as fast as data centers wanted.

    “I am unaware of any individuals bringing forward any efforts at this time to start a next speed project after the 400GbE standard is completed,” said John D’Amborisa, chairman of IEEE 802.3 New Ethernet Applications, an ad hoc group used for industry consensus building.

    Nevertheless, “the Ethernet Alliance [trade group] has recognized 800GbE as a logical next step after 400GbE on its Ethernet roadmap, so there is no real surprise at individuals calling out the need for this standard,” said D'Ambrosia who is also chairman of the Ethernet Alliance and several IEEE past Ethernet standards efforts.

    Meanwhile the Optical Internetworking Forum (OIF) is already hammering out multiple interfaces for 112G serial links. Their efforts come as demos of 56G serdes using PAM-4 are now emerging at events such as last week’s DesignCon.

    The OIF also is working on a Flexible Ethernet standard. It aims to cover a wide range of data rates up to 100G in chips arriving next year with a 400G follow-on in the works.

    Bechtolsheim said the FlexE standard will be too expensive for the mainstream switch chips data centers need. “It’s a good idea, the only problem is it’s a very low volume market,” he said, characterizing it as “a very expensive chip for very small market.”

    The FlexE work “was broadly supported by industry leaders, including end users and their system and chip suppliers, with further developments continuing as a result of that same broad industry support,” said Scott Irwin, protocol vice chair of the OIF’s PLL Working Group. “FlexE 1.0 is based on standard 100GE PHYs, requiring only incremental logic inside a 100GE switch chip,” he noted, suggesting it would not carry high incremental costs.

    Beyond the 800G generation, Web giants may need to move to optics on the board, an initiative Microsoft’s Azure cloud group has been championing since early 2015. The Consortium for On-Board Optics aims to give and update on its progress at the Optical Fibers Conference in March where vendors wil stage a demo of 400G systems. On-board optics maybe required as early as 2021, said Bechtolsheim.

    The good news, according to Bechtolsheim, is merchant chip vendors led by Broadcom have reduced the need for switching ASICs and FPGAs. Although pursuing Moore’s law is becoming more costly and complex, cost per bit of bandwidth continues to decline, he said.

    “There’s still a factor of 100 in transistor density ahead, and networking is coming from far behind, with 28nm chips shipping today and this year 16nm ones coming,” he said.

    Although Broadcom dominates the switch chip market, Arista started using Xpliant chips from Cavium Networks this year, said analyst Wheeler. Barefoot Networks has first silicon on another switch and startup Innovium is working on a switch chip, he added.

    The chips dominate switch systems such that “all my work is software--95% of my headcount is in software engineering,” Bechtolsheim said.

  2. #2
    WHT-BR Top Member
    Data de Ingresso
    Dec 2010

    Bechtolsheim: Networking Standards Move Too Slowly for Cloud Giants

    Craig Matsumoto
    February 8, 2017

    The rise of the cloud titans has disrupted the comfortable standards cycles of the networking world and is — or at least should be — forcing the ecosystem to move faster, Arista Co-Founder Andy Bechtolsheim said during the opening talk at the Linley Cloud Hardware Conference today.

    “Ten years ago, there was a single company dominating the space, and — let’s just say they didn’t have a motivation to go faster,” he said. “In the new world, it’s a fundamental change in the competitive environment.”

    Of course, that “single company” was Cisco — a former employer and current competitor of his. But all systems vendors are wrestling with the pace of the cloud world, he said, partly because they all rely on components vendors who, themselves, are having trouble adjusting to cloud speed.

    That’s why Arista is trying to start a grass-roots standards effort for 800 Gb/s optics, which Bechtolsheim plugged briefly during his talk.

    An IEEE standard for 800 Gb/s would take at least three years, and Bechtolsheim doesn’t believe cloud providers will wait that long. In fact, the 400 Gb/s standard isn’t likely to be finalized until 2018, at which point it will have been in the works for four years, he said.

    “We don’t have time for this. The design activity for 800 Gb/s has to start this calendar year.”

    A Faster World

    One reason standards have to speed up is because cloud vendors adopt new technologies more abruptly than do traditional customers such as telecom providers, Bechtolsheim said.

    A new speed grade for Ethernet — 10 Gb/s, say — used to be accompanied by a years-long transition in which customers would apply the new, expensive technology sparingly while waiting for prices to fall, he said.

    Cloud vendors behave differently. “They will not buy the new thing unless it is cheaper than the old thing,” he said. Moreover, “the moment it’s cheaper than the old stuff, nobody wants the old stuff any more.”

    The result is that a years-long transition between Ethernet generations has shrunk to about six months.

    In the case of 100-Gb/s Ethernet, that led to some surprises. Prices for 100-Gb/s optical interfaces fell more quickly than expected, causing demand to surge and inventories to run dry. Bechtolsheim estimated the number of ports shipped at less than 1 million last year, 5 million this year, and likely 10 million next year.

    “The optics industry wasn’t ready for that,” he said.

    That Latest Chip

    Things are similarly changing at the chip level, thanks to emerging high-end options for merchant silicon, Bechtolsheim said. In addition to market leader Broadcom, competitive chips from vendors such as Cavium and Barefoot Networks are now reaching the market.

    As cloud vendors consider switches based on those chips, it creates another source of pressure for systems vendors, such as Cisco, that design their own semiconductors. Bechtolsheim founded Arista partly on the thesis that off-the-shelf chips could produce a competitive high-end switch. He returned to that theme during his talk today.

    Cloud vendors “want that latest chip,” he said. “They’re not going to wait for the legacy company to tell them it’s all right to buy it. They’ve got to buy it as soon as it works.”
    Última edição por 5ms; 09-02-2017 às 11:44.

Permissões de Postagem

  • Você não pode iniciar novos tópicos
  • Você não pode enviar respostas
  • Você não pode enviar anexos
  • Você não pode editar suas mensagens