inphi moves big data faster · 1/13/2021  · • electrical input: 8 x 26.5625 gbaud pam 4, ieee...

22
R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 1 Inphi Copyright Inphi Moves Big Data Faster Next-Gen Data Center Interconnects: The Race to 800G Radhakrishnan Nagarajan, Ilya Lyubomirsky Jan 13, 2021 COBO Webcast

Upload: others

Post on 25-Jan-2021

10 views

Category:

Documents


0 download

TRANSCRIPT

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 1Inphi Copyright

    Inphi Moves Big Data Faster

    Next-Gen Data Center Interconnects: The Race to 800G

    Radhakrishnan Nagarajan, Ilya Lyubomirsky

    Jan 13, 2021COBO Webcast

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 2Inphi Copyright

    Data center interconnect evolution ➔ hierarchical to flatter architectures

    • Elimination of layers

    • Very high-speed interconnects between layers

    • Elimination/minimization of over-subscription

    • Datacenters distributed over large geographic distances

    - Source: Microsoft

    Interconnect-rich

    Machine to machine traffic dominates

    High radix connections

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 3Inphi Copyright

    Why do we care about the datacenter radix?

    Switch Generation Radix = 32 Radix = 64 Radix = 128

    12.8T 400G 200G 100G

    25.6T 800G 400G 200G

    51.2T 1.6T 800G 400G

    Network flattening

    Optical

    interconnects

    -Source: Facebook

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 4Inphi Copyright

    Ubiquitous 2x speed scaling

    60Gbaud, QAM 16, 1λ

    53Gbaud, PAM4, 4 λ’s

    2x λ’s2x baud rate

    Higher order

    modulation format

    … analog bandwidth … power, size, cost, not scalable

    … complexity, SNR

    400G, 4λ’s ➔ 800G, 8λ’s

    QAM 256PAM 16

    Engineering trade-offs

    QAM 64, 90Gbaud

    400GPAM 6, 90Gbaud

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 5Inphi Copyright

    Technology choices in data center interconnects

    IEEE 802.3

    PSM4, CWDM4

    NRZ

    25G/λ,CWDM

    PAM4

    Coherent

    100G to 600GDWDM

    Coherent

    QPSK/8QAM/16QAM

    100G/200G/400GDWDM

    Coherent

    QPSK/8QAM

    100G/200G/300GDWDM

    Inside DC DCI-Campus/Edge DCI-Metro DCI-Long Haul

    View at 100G

    IEEE 802.3

    LR4

    NRZ

    25G/λCWDM

    IEEE 802.3

    DR4, FR4

    PAM4

    100G/λCWDM

    Coherent, OIF/IEEE/ITU-T

    QPSK/8QAM/16QAM (400G ZR)

    100G/200G/400G Pluggable ModuleDWDM

    View at 400G

    Distance (km)0.1 2.0 10 100 600 10000

    IEEE 802.3

    DR4, FR4

    PAM4

    200G/λCWDM

    Coherent, OIF

    QPSK/8QAM/16QAM (800G ZR)

    100G/200G/400G Pluggable ModuleDWDM

    View at 800G

    CWDM LR

    coherent

    Distance

    optimized,

    low power

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 6Inphi Copyright

    Inside data center PAM 4: 400G evolution to 800G

    400G ➔ 4 λ 800G ➔ 8 λ 800G ➔ 4 λ

    Host side interface 50G PAM4 100G PAM4 100G PAM4

    Number of host side high speed lanes 8 8 8

    Line side interface 100G PAM4 100G PAM4 200G PAM4

    DSP CMOS node 7nm 7nm / 5nm 5nm

    PAM4 DSP die area 1.0 2.0 (7nm) 1.X

    PAM4 DSP power 1.0 2.0 (7nm) 1.4

    Other optics, TIA, Driver power 1.0 2.0 1.X

    Optics + electronics BW requirements ~ 40GHz ~40GHz ~65GHz

    Optical laser source power 1.0 2.0 1.X

    Smallest pluggable form factor QSFP-DD/OSFP QSFP-DD/OSFP QSFP-DD/OSFP

    200G/λ is necessary to 1.6T modules

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 7Inphi Copyright

    Inside data center: Coherent vs PAM

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 8Inphi Copyright

    Pluggable module form factors for data center interconnects

    QSFP-28 QSFP-DD 2

    OSFP

    CFP2

    • QSFP-DD maintains the same width as the QSFP-28 and is backward compatible.

    • OSFP is a slightly wider form factor.

    • A 1RU (1.75” height) switch chassis, in a 19” (width) rack, will accommodate 32 modules of QSFP, QSFP-DD and OSFP form factors.

    100G PAM 4 DCI module

    400G QAM 16 DCI module

    CFP2 module for size comparison

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 9Inphi Copyright

    High speed optical interconnects enabled by silicon photonics

    High speed Mach Zehnder modulator High speed Ge photodetector

    Polarization beam splitter • Performance of key components in the integrated Silicon

    photonics chip across C band.

    • High speed Mach Zehnder modulator

    • High speed Ge photodetector

    • Low loss, high extinction ration polarization beam splitter.

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 10Inphi Copyright

    Between data centers: silicon photonics based 400G ZR coherent modules

    • QSFP-DD module ➔ similar form factor to DR4 and FR4 modules

    • Advanced 7nm CMOS DSP, Canopus

    • Advanced silicon photonics node from Bi-CMOS foundry

    • Electrical Input: 8 x 26.5625 Gbaud PAM 4, IEEE 802.3bs

    • Optical Output: 1 x 59.8 Gbaud QAM 16, 1 λ

    • 80km-120km single span reach

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 11Inphi Copyright

    Between data centers: silicon photonics based 400G ZR performance at 80km

    • Polarization scrambled at 50krad/s

    • < 0.5dB polarization penalty

    400G switch

    Multi-channel

    spectrum

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 12Inphi Copyright

    200G/λ: Modulation,

    Equalization and FEC

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 13Inphi Copyright

    Optical standards for data center interconnect

    Standard ModulationBaud Rate

    (Gbaud)

    Wavelength

    (nm)

    Distance

    (m)FEC

    FEC NCG

    (dB)

    FEC Latency

    (ns)

    400GBASE-SR8 PAM4 26.56 850 100 RS(544,514) 6.8 100

    400GBASE-FR8 PAM4 26.56 1300-LWDM 2,000 RS(544,514) 6.8 100

    400GBASE-LR8 PAM4 26.56 1300-LWDM 10,000 RS(544,514) 6.8 100

    400GBASE-ER8* PAM4 26.56 1300-LWDM 40,000 RS(544,514) 6.8 100

    400GBASE-SR4* PAM4 53.125 850 50 RS(544,514) 6.8 100

    400GBASE-DR4 PAM4 53.125 1300 500 RS(544,514) 6.8 100

    400GBASE-FR4 PAM4 53.125 1300-CWDM 2,000 RS(544,514) 6.8 100

    400GBASE-LR4-6 PAM4 53.125 1300-CWDM 6,000 RS(544,514) 6.8 100

    400GBASE-LR4-10* PAM4 53.125 1300-CWDM 10,000 RS(544,514) 6.8 100

    800GBASE** PAM4 (6) ~112-114 1300-CWDM ~1-2kmHigher Gain

    FEC~9-10 < 300

    * Under development

    ** IEEE 802.3 Beyond 400GbE Study Group starting work in Jan. 2021

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 14Inphi Copyright

    Source: 800G Pluggable MSA, White Paper, “Enabling the Next Generation of Cloud & AI using 800 Gb/s Optical Modules”

    ▪ Transmitter analog bandwidth limitations and electrical reflections (see example below)

    ▪ Laser RIN noise (level dependent noise) - impacts PAM6 more than PAM4

    ▪ Fiber chromatic dispersion – penalty scales quadratically with baud rate

    ▪ Polarization mode dispersion, first order effect due to differential group delay (DGD)

    ▪ Receiver TIA noise, nonlinearity, and analog bandwidth limitations

    Driver Laser/Modulator PD/TIA ADCDAC

    S21 (dB)

    Key challenges for 200G/λ PAM optical systems

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 15Inphi Copyright

    PAM Rx DSP architectures for beyond 400GbE

    ADC 10-30 tap FFE1 tap

    DFE

    FEC

    Baud Rate

    1/T

    Analog Digital

    PLL DCO TEDLF

    Err Gen

    LMS

    MLSD

    Mux

    LMS

    AFE

    Slicer

    1. High Speed ADC

    enables

    DSP Architectures

    2. FFE, DFE, and

    MLSD for stronger

    EQ

    3. Leverage DSP soft

    information for

    higher coding gain

    FEC

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 16Inphi Copyright

    Concatenated Codes: ADC Enables Soft Decision Inner Code

    Source: Frank Kschischang, “Introduction to Forward Error Correction,” OFC Short Course, 2018

    KP4

    (e.g., Host FEC)

    KP4

    (e.g., Host FEC)Inside Module Inside Module

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 17Inphi Copyright

    Low latency higher gain FEC simulation results

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 18Inphi Copyright

    Baud Rate:

    113Gbaud, PAM 4

    90Gbaud, PAM 6

    Optical System:

    RX OMA = -4 dBm

    Laser RIN = -145 dB/Hz

    ER = 5 dB

    TIA NEP = 12 pA/sqrt(Hz)

    FEC Limit

    Comparison of modulation formats for 200G/λ

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 19Inphi Copyright

    >FEC threshold >FEC threshold

    PAM4

    113.3GBaud

    PAM6

    90.7GBaud

    PAM4 vs PAM6 model with realistic component models

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 20Inphi Copyright

    Tx chirp

    FR4 2km

    200G/λ PAM4 chromatic dispersion penalty

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 21Inphi Copyright

    Conclusions

    ▪ Next generation 800G data center optical interconnects should be designed considering the

    unique application requirements of data center operators.

    ▪ The choice of 200G/λ modulation format, DSP architecture, and FEC scheme requires careful

    analysis on the engineering tradeoffs for analog component technology and optical channel

    specifications.

    ▪ Optical system simulations show PAM4 200G/λ is a feasible technology for 800G-FR4/DR4,

    synergistic with low latency AI, break out, and copper (DAC) applications.

    ▪ 800G coherent technology may be most useful in the longer reach applications such as 800G

    for 40km, where coherent dispersion tolerance and lower fiber loss at 1550 nm are the key

    advantages.

  • R. Nagarajan, Ilya Lyubomirsky, “Next-Gen Data Center Interconnects: The Race to 800G” 22Inphi Copyright

    Inphi Moves Big Data Faster