Need for Speed: Beyond 100GbE Moderator: Scott Kipp, President of Ethernet Alliance, Principle Engineer, Brocade Panelist #1: Alan Weckel, Vice President, Dell’Oro group Panelist #2: Dr. Jeffery J. Maki, Distinguished Engineer, Juniper Panelist #3: Dr. Gordon Brebner, Distinguished Engineer, Xilinx © 2013 Ethernet Alliance 1 Agenda Introductions: Scott Kipp, Moderator Panelist #1: Alan Weckel, Panelist #2: Dr. Jeffery J. Maki, Stepping Stones to Terabit-Class Ethernet Panelist #3: Dr. Gordon Brebner, 10, 40 and 100GbE Deployments in the Data Center Technology Advances in 400GbE Components Q&A 2:40 – Live Broadcast from IEEE 802.3 Meeting in Orlando from John D’Ambrosia Update on 400GbE Call For Interest © 2013 Ethernet Alliance © 2012 Ethernet Alliance 2 Disclaimer The views WE ARE expressing in this presentation are our own personal views and should not be considered the views or positions of the Ethernet Alliance. © 2013 Ethernet Alliance 3 Bandwidth Growth Broadband 2010- 7Mbps 2015 – 28 Mbps 15B Devices In 2015 Increased # of More Users Increased + Access Devices Rates and Methods More Internet Users + Increased Services Key Growth Factors Speed = Bandwidth Explosion Increasing Everywhere More Rich Media Content 2010- 1 Minute video 2015 – 2 hour HDTV Movie 3B Users In 2015 Source: nowell_01_0911.pdf citing Cisco Visual Networking Index (VNI) Global IP Traffic Forecast, 2010–2015, http://www.ieee802.org/3/ad_hoc/bwa/public/sep11/nowell_01_0911.pdf © 2013 Ethernet Alliance 4 Bandwidth Growth Vs Ethernet Speeds IP Traffic is growing ~ 30%/year If 400GbE is released in 2016, Ethernet speeds will grow at about 26%/year Ethernet Speed (Gb/s) Internet traffic normalized to 100 in 2010 1600 1400 1200 1000 800 600 Internet traffic would grow ~10X by 2019 at 30%/year Ethernet speeds to grow 4X by 2016 at 26%/year Ethernet Speed Internet Traffic 400 200 0 © 2013 Ethernet Alliance 5 Data Rate and Line Rate (b/s) Ethernet Optical Modules CFP CFP2 100G 300 Pin MSA XENPAK XPAK X2 CFP4 100GbE XFP CXP 40GbE 40G QSFP28 QSFP+ 10GbE 10G SFP+ GBIC SFP Key: Ethernet Standard Released Module Form Factor Released GbE 1G 1995 2000 2005 2010 2015 Standard Completed © 2013 Ethernet Alliance 6 Data Rate and Line Rate (b/s) Ethernet Speeds 2010-2025 If Ethernet line rates doubles the line rate every 3 years at 26% CAGR, then 400GbE would come out in 2016 and TbE would come out in 2020. Something will have to change. 1.6TbE 16X100G TbE 10X100G 1T 400G 100G 40G 10G 400GbE 16X25G 100GbE 10X10G 400GbE 8X50G 100GbE 4X25G 100GbE 1X100G 2010 8X50G 4x25G 2015 16x25G 2020 Ethernet Electrical Interfaces Hollow Symbols = predictions Stretched Symbols = Time Tolerance nX100G 40GbE 4X10G 4x10G 10X10G 400GbE 4X100G Key: Ethernet Speeds 2025 Standard Completed © 2013 Ethernet Alliance 7 Ethernet Success Ethernet has been extremely successful at lowering the price/bit of bandwidth If the cost of a new speed/technology is too high, then it is not widely deployed Technology needs to be ripe for picking 400GbE is ripe with 100GbE technology TbE isn’t ripe and a revolutionary breakthrough would be needed to get it before 2020 This panel will look at how high speeds of Ethernet are being deployed and the technology that is leading to the next generation of Ethernet © 2013 Ethernet Alliance 8 10, 40 and 100GbE Deployments in the Data Center Alan Weckel Vice President, Data Center Research Dell’Oro Group © 2013 Ethernet Alliance 9 Introduction Progress on server migration from 1 GbE to 10 GbE 10G Base-T update Data center networking market update 40 GbE and 100 GbE market forecasts © 2013 Ethernet Alliance 10 Overview Dell’Oro Group is a market research firm that has been tracking the Ethernet Switch and Routing markets on a quarterly basis since 1996 We also track the SAN market, Optical market, and most Telecom equipment markets We produce quarterly market share reports that include port shipments as well as market forecasts © 2013 Ethernet Alliance 11 700 350 © 2013 Ethernet Alliance 20 16 20 15 20 14 20 13 20 12 20 11 20 10 0 20 09 Petabytes per Second Shipped per Year Data Center Bandwidth Shipping – Ethernet Switching 12 1 GbE 40 GbE 10 GbE 100% 75% 50% 25% © 2013 Ethernet Alliance 20 16 20 15 20 14 20 13 20 12 20 11 20 10 0% 20 09 Percent of Server Shipments Switch Attach Rate on Servers 13 180 10G Base-T controller and adapter ports 90 10G Base-T switch ports © 2013 Ethernet Alliance 12 4Q 12 3Q 12 2Q 12 1Q 11 4Q 11 3Q 11 2Q 11 0 1Q Port Shipments in Thousands Data Center Port Shipments – 10 G Base-T Port Shipments 14 50 1 GbE 10 GbE 40 GbE 100 GbE 25 © 2013 Ethernet Alliance 20 16 20 15 20 14 20 13 20 12 20 11 20 10 0 20 09 Port Shipments in Millions Data Center Port Shipments – Ethernet Switching 15 6 40 GbE 100 GbE 3 © 2013 Ethernet Alliance 20 16 20 15 20 14 20 13 20 12 20 11 20 10 0 20 09 Port Shipments in Millions Data Center Port Shipments – Ethernet Switching 16 Summary Ethernet Switches will be responsible for the majority of 40 GbE and 100 GbE port shipments over the next five years Form-factor and cost driving 40 GbE over 100 GbE 10 GbE server access transition is key to higher speed adoption © 2013 Ethernet Alliance 17 Stepping Stones to Terabit-Class Ethernet: Electrical Interface Rates and Optics Technology Reuse Jeffery J. Maki Distinguished Engineer, Optical Juniper Networks, Inc. © 2013 Ethernet Alliance 18 100G © 2013 Ethernet Alliance 19 CFP, CFP2 and CFP4 for SMF or MMF Applications CFP MSA Form Factors: http://www.cfp-msa.org/ CFP4 CFP2 CFP Optical Connector • LC Duplex (depicted) Courtesy of TE Connectivity • MPO © 2013 Ethernet Alliance 20 Module Electrical Lane Capability CFP4 CFP2 CFP 12x10G electrical lanes 10x10G or 8x25G electrical lanes 4x25G electrical lanes CAUI for 10x10G CPPI & CAUI for 10x10G CAUI-4 for 4x25G CAUI-4 for 4x25G © 2013 Ethernet Alliance 21 CFP, CFP2, and CFP4 for 100G Ethernet SMF PMD Gear Box Transmit side only depicted. LAN WDM 1295.56 nm 1300.05 nm 1304.58 nm 1309.14 nm Gear Box LAN WDM 1295.56 nm 1300.05 nm 1304.58 nm 1309.14 nm 4 λ on LAN WDM CFP Current Options • Up to 10 km: 100GBASE-LR4 • Up to 40 km: 100GBASE-ER4 CFP2 CFP4 © 2013 Ethernet Alliance 22 400G © 2013 Ethernet Alliance 23 Projection of Form Factor Evolution to 400G speculation CD-CFP4 16x25G electrical lanes CD-CFP2 CD-CFP CFP4 CFP2 CFP 4x100G 8x50G electrical electrical lanes lanes CFP4 CFP4 CFP4 CFP4 Roman Numerals XL = 40 C = 100 CD = 400 400G defensible 100G © 2013 Ethernet Alliance 24 Likely MSA Activity CFP MSA http://www.cfp-msa.org/ CD-CFP: Current CFP needs revamping to support 16 x 25G CD-CFP2: Current CFP2 is ready for 8 x 50G CD-CFP4: Unclear New CDFP MSA http://www.cdfp-msa.org/ High-density form factor supporting 16 x 25G From slide 26 of http://www.ieee802.org/3/cfi/0313_1/CFI_01_0313.pdf © 2013 Ethernet Alliance 25 400G Optics Requirements First-generation transceivers have to be implementable that meet and eventually do better than these requirements Size (Width): 82 mm (CFP width, ~4 x CFP4) Cost: 4 x CFP4 Power: 24 W (4 x 6 W power profile of CFP4) Improved bandwidth density transceivers will need higher rate electrical-lane technology 50G 100G © 2013 Ethernet Alliance 26 How 400G Ethernet Can Leverage 100G Ethernet 100G Ethernet up to 10 km CFP4-LR4 Duplex Single-Mode Fiber Infrastructure CFP4-LR4 400G Ethernet up to 10 km Parallel Single-Mode Fiber Infrastructure CFP4-LR4 CFP4-LR4 Only 8 Fibers Used CFP4-LR4 CFP4-LR4 CFP4-LR4 CFP4-LR4 CFP4-LR4 CFP4-LR4 © 2013 Ethernet Alliance 27 Possible SMF Ethernet Road Map: 100G, 400G, 1.6T Early Adopter 400G Mature 400G 4 x 100GBASE-LR4 or “400GBASE-PSM4” 400GBASE-??? CFP4(LC) CD-CFP2(LC) CD-CFP4(LC) CFP4(LC) CD-CFP(MPO) CD-CFP2(MPO) 4 x 400GBASE-??? or “1600GBASE-PSM4” CD-CFP4(LC) CD-CFP4(LC) CFP4(LC) CFP4(LC) Early Adopter 1.6T CD-CFP4(LC) CD-CFP4(LC) Parallel Single Mode, 4 Lanes (PSM4) 4, Tx Fibers and 4, Rx Fibers 1x12 MPO Connector (High-Density 100GE) © 2013 Ethernet Alliance 28 Early Adopter 400G using SMF Structured Cabling Parallel SMF: “400GBASE-PSM4” Technology Reuse: 4 x 100GBASE-LR4 Courtesy of Commscope © 2013 Ethernet Alliance 29 Early Adopter 400G using MMF Structured Cabling Courtesy of Commscope Technology Reuse: 4 x 100GBASE-SR4 Parallel MMF: “400GBASE-SR16” Parallel Multi-Mode • 100GBASE-SR4, 4 x 25G optical lanes: 4, Tx Fibers and 4, Rx Fibers using 1x12 MPO • “400GBASE-SR16”, 16 x 25G optical lanes: 16, TX Fibers and 16, Rx Fibers using 2x16 MPO © 2013 Ethernet Alliance 30 MMF Breakout Cables— Enabling 400G Adoption 2 x 16 MPO 1 x 12 (8 used) MPO 2 x 16 MMF MT ferrule 1 x 12 (8 used) MPO 1 x 12 (8 used) MPO 1 x 12 (8 used) MPO © 2013 Ethernet Alliance Courtesy of USConec 31 100G Can Build 400G at the Cost of 4 x 100G Technology Reuse: 4 x 100GBASE-LR4 Parallel SMF: “400GBASE-PSM4” Technology Reuse: 4 x 100GBASE-SR4 © 2013 Ethernet Alliance Parallel MMF: “400GBASE-SR16” 32 Ethernet PMD Maturity & Possible Obsolescence Early Adopter PMD Parallel Fiber, SMF or MMF Leverage of mature PMD from previous speed of Ethernet Planned obsolescence Implementation (with MPO connector) persists as high-density support of previous speed of Ethernet (e.g., 4 x 100G) Mature PMD SMF: Duplex SMF cabling (e.g., with LC duplex connector) MMF: Lower fiber count MMF cabling © 2013 Ethernet Alliance 33 SMF Density Road Map 16 CD-CFP4(LC) (mature) Front-Panel Bandwidth Density (Relative) 8 (early adopter) CD-CFP2(MPO) CD-CFP2(MPO) CD-CFP2(LC) (mature) (early adopter) (mature) 4 CFP4(LC) 2 1 CFP2(LC) CFP(LC) 100G 4x CD-CFP4(LC) 4 x CFP4(LC) or CD-CFP(MPO) (early adopter) Port Bandwidth 400G © 2013 Ethernet Alliance 1.6T 34 Summary Form-factor road map for bandwidth evolution Early adopter 400G Ethernet by reusing 100G module and parallel cabling, SMF or MMF Need for a new, 2 x 16 MMF MT ferrule Possible common module for 400G Ethernet and high-density (4-port) 100G Ethernet Need for new electrical interface definitions supporting lane rates at 50G 100G © 2013 Ethernet Alliance 35 Technology Advances in 400GbE Components Gordon Brebner Distinguished Engineer Xilinx, Inc. © 2013 Ethernet Alliance 36 400GbE PCS/MAC Expect first: 16 PCS lanes, each at 25.78125 Gbps Glueless interface to optics Possible re-use of the 802.3ba PCS Other options possible for PCS, maybe native FEC Later: 8 lanes, each at 51.56Gbps Or 4 lanes with 2 bits/symbol at 56Gbaud (e.g. PAM4) Packet size 64 bytes to 9600 bytes Use 100GbE building blocks where possible © 2013 Ethernet Alliance 37 Silicon technology Technology nodes (silicon feature size) Application-Specific Integrated Circuit (ASIC) 130nm, 65nm, 40nm, 28/32nm, 20/22nm, 14/16nm Fixed chip Increasingly expensive: need high volumes Best suited to post-standardization Ethernet Field Programmable Gate Array (FPGA) Programmable logic chip Suitable for prototyping and medium volumes Best choice for pre-standardization Ethernet © 2013 Ethernet Alliance 38 400GbE line/system bridge Wide parallel data path between blocks CDFP or 4xCFP4 16 x 25G SERDES 400GbE 400GbE PMA/PCS MAC Bridge logic 40 x 12.5G or 48 x 10G SERDES 500G Interlaken Optical ASIC or FPGA chip Line side System side © 2013 Ethernet Alliance 39 MAC rate = Width x Clock 400 Gbps and 1 Tbps Ethernet MAC options MAC rate Silicon node Technology Data path width Clock frequency 100 Gbps 45, 40nm ASIC 160 bits 644 MHz 100 Gbps 45, 40nm FPGA 512 bits 195 MHz 400 Gbps 28, 20nm ASIC 400 bits 1 GHz 400 Gbps 28, 20nm FPGA 1024 bits 1536 bits 400 MHz 267 MHz 1 Tbps 20, 14nm ASIC 1024 bits 1 GHz 1 Tbps 20, 14nm FPGA 2048 bits 2560 bits 488 MHz 400 MHz © 2013 Ethernet Alliance 40 Multiple Packets/Word Bus width Max packets Max EOPs 512 2 1 1024 3 2 1536 4 3 512 * n n+1 n Up to 512-bit, only one packet completed Just need to deal with EOP then SOP in word Beyond 512-bit, multiple packets completed Need to add parallel packet processing Must deal with varying EOP and SOP positions © 2013 Ethernet Alliance 41 400GbE CRC Example All Ethernet packets carry Cyclic Redundancy Code (CRC) for error detection Computed using CRC-32 polynomial Critical function within Ethernet MAC Requirements Computed at line rate Deal with multiple packets in wide data path Economical with silicon resources © 2013 Ethernet Alliance 42 400GbE CRC Prototype Xilinx Labs research project Modular: built out of 512-bit 100G units Computes multiple CRCs per data path word Targeting 28nm FPGA (Xilinx Virtex-7 FPGAs) 512-bit unit CRC results combined to get final CRC results N-bit data path partitioned into 512-bit sections © 2013 Ethernet Alliance 43 400GbE CRC Prototype Results: Data bus word size 1024-bit 1536-bit 2048-bit Max clock frequency (MHz) 400 381 326 Maximum line rate (Gbps) 409 585 668 Latency (ns) 17.5 18.4 21.5 FPGA resources (slices) 2,888 4,410 5,719 1024-bit width is feasible for 400GbE Other widths: Less challenging clock frequencies Demonstrate scalability beyond 400GbE © 2013 Ethernet Alliance 44 Conclusions Can anticipate 400GbE PCS/MAC standard Ever-increasing rates mean ever-wider internal data path width in electronics Leading to multiple packets per data word Possible to prototype pre-standard PCS/MAC using today’s FPGA technology Demonstrated modular Ethernet CRC block based on 100GbE units Silicon resource scales linearly with line rate © 2013 Ethernet Alliance 45