1 of 13

OCP External PCIe(NVMe & CXL) Connectivity Workstream

  • OCP Server Project
  • PCIe External Connectivity Workstream
  • Dec 12th, 2022

​

Connect. Collaborate. Accelerate.

Seagate Internal

2 of 13

Extended PCIe Cable Market Needs Timeline

PCIe Gen5

PCIe Gen6 🡪

2022

2023

2024

2025

2026

2027

Aug-22

Fabric Optimized

CXL3.0 Specification Release

2H-25

Early CXL1.1/2.0 PCIe-Gen5 Pooled/Shared Appliances

2H-27

Mature CXL3.0/3.1 PCIe-Gen6 Pooled/Shared Emerging Memory Appliance Deployment

2H-26

First PCIe-Gen5 Host attached NVMe HDD Mass Storage JBOD

1H-24

Early Pooled DRAM PCIe-Gen5 CXL1.1/2.0 Deployments

1H-23

NVMe HDD EDU2 Availability from Seagate

2H-22

NVMe HDDs Engineering Demo Units(EDU) Demonstrated

Years ago

PCIe-Gen4 NVMe SSDs are deployed in JBOFs

Nov-2020

Fabric Capable CXL2.0 Specification Release

1H-26

PCIe-Gen3 NVMe Possible HDD CTUs Availability

2H-24

NVMe HDD Possible CDU Availability

1H-23

Expected CXL3.1 Specification Release

  • PCIe-Gen5 NVMe/CXL x4, x8 & x16
  • PCIe-Gen5 NVMe/CXL x4, x8 & x16

2H-23

PCIe-Gen5 Host attached NVMe SSD Primary Storage JBOF

1H-27

PCIe-Gen6 Host attached CXL/NVMe SSD Primary JBOF

2H-27

NVMe attached SSD & HDD to replace SATA JBODs

CXL3.0/3.1

Gen6 Hosts

CXL2.0+

Gen5 Hosts

CXL1.1

Gen5 Hosts

🡪

Seagate Internal

3 of 13

NVMe Connectivity Scenarios & Requirements

NVMe Connectivity for DAC, AEC and AOC:

  • Compute to Storage x4 & x8 & x16 PCIe-Gen5
  • Accelerator to Storage x4 & x8 & x16 PCIe-Gen5
  • Storage Expansion x4 & x8 & x16 PCIe-Gen5
    • Distance 0.5 to 5 Meter distance requirements (Wen Mei Hwu nVidia)
    • Cable Management Arm(CMA) Restrictions (Barrett Edwards - WDC)
      • Bend Radius Restrictions for CMA 320-deg ?
      • Extra length needed for CMA on each end 1.2 to 1.5M ?
      • Consider Interoperability between Connectors of different width
        • Support for different width and connectors types on ether sides of the cable
        • Support multiple side-band signals for bifurcation
          • Support either bifurcated and non-bifurcate cables
          • Support for SRIS/SRNS specific cables
      • DAC, Gen5 32Gb/s, BER 10-12, 28 to 30 gauge Cable, Bend-Radius Budget?
        • With Passive CEM card RP/EP 8”> x >4” from connector on main PCB on both ends 🡪 0.5M
        • With Active CEM card RP/EP 4”> x >2” from connector on main PCB on both ends 🡪 0.5M to 1M
      • AEC, Gen5 32Gb/s, BER 10-12, 32 gauge Cable, 36-DBL each end
        • Bend-Radius Budget 5x outside diameter of the cable
        • Sideband signals of SmBus/I2C/I3C are localized to the connector and will not connect between the two ends of the cable
        • REFCLK & PERST are not needed due to the proposed use of SRIS/SRNS
        • Re-Timer Based Solutions 🡪 1M to 3M @ 32-gauge
          • Power budget for x4, x8 & 16 with 1.15W per lane including DC/DC deficiency
          • For a x4 cable 5.2mm Diameter (Don from Credo) add cable bends
          • For a x8 cable 6.8mm Diameter (Don from Credo) add cable bends – x8 Retimer package size
          • For a x16 cable we use two of the x8 cables add cable bends – x16 Retimer package size 8.90 mm x 22.80 mm
        • Re-Driver Based Solutions 🡪 1M to 2M? @ 32-gauge – Alvin Cox Re-driver add significant Jitter and could prevent us from any improvements over DAC cable.
        • Active device on both-ends are in most cases
          • Active device is added on one end of the other on the Root-port or Endpoint card if active device in not needed on both ends of the wire
      • AOC or other Optical NPO/CPO, Gen5 32Gb/s , BER 10-12 🡪 1M to 10M
        • Electrical specification – Chip to module distance
          • Long-Reach Full Standard PCIe Gen5 36 DBL budget – Required for most Active cable solutions in order to interoperate with standard PCIe Gen5 transceivers
          • Medium to Short Reach DBL budget ?
          • Very-Short-Reach DBL budget ?
        • Multi-Mode or Single-Mode fiber cables can be used as long as they are cost effective
        • Bend-Radius Budget 10x optical fiber diameter ~30mm
        • Sideband signals of SmBus/I2C/I3C are localized to the connector and will not connect between the two ends of the cable
        • REFCLK & PERST are not needed due to the proposed use of SRIS/SRNS

GPUs

SSDs

HDDs

DRAM

EM

SCM

Compute Nodes

CXL

NVMe

CXL =

NVMe =

CXL Switch

PCIe/NVMe Switch

DPUs/RNICs/NICS …etc

Seagate Internal

4 of 13

CXL Connectivity Scenarios & Requirements

CXL Connectivity for DAC, AEC and AOC:

  • Compute Cluster x8 & x16 PCIe-Gen5
  • Compute to Accelerator x8 & x16 PCIe-Gen5
  • Compute to Memory x8 & x16 PCIe-Gen5
  • Accelerator Cluster x8 & x16 PCIe-Gen5
  • Accelerator to Memory x8 & x16 PCIe-Gen5
    • Distance 0.5 to 5 Meter distance requirements
      • Bend radius
      • Consider Interoperability between Connectors of different width
        • Support for different width and connectors types on ether sides of the cable
        • Support multiple side-band signals for bifurcation
          • Support either bifurcated and non-bifurcate cables
          • Support for SRIS/SRNS specific cables
      • DAC, Gen5 32Gb/s, BER 10-12, 8-DBL? each end, 28 to 30-gauge Cable, Bend-Radius Budget 🡪 0.5M to 1M
        • With Active RP/EP 6”> x >2” from connector on main PCB on both ends
      • AEC, Gen5 32Gb/s, BER 10-12, 16-DBL each end, 32-gauge Cable, Bend-Radius Budget 🡪 2M to 4M
        • Re-Timer Based?
        • Re-Driver Based?
        • Power Budget?
        • Additional Pins Needed?
        • Active device on both-ends ?
        • Active device on one-end?
      • AOC, Gen5 32Gb/s , BER 10-12, 16-DBL each end, Multi-Mode Cable, Bend-Radius Budget 🡪 5M to 10M

GPUs

SSDs

HDDs

DRAM

EM

SCM

Compute Nodes

CXL

NVMe

CXL =

NVMe =

CXL Switch

PCIe/NVMe Switch

DPUs/RNICs/NICS …etc

Seagate Internal

5 of 13

OCP Workstream Charter 

  • Research and document the future Compute, Storage, Accelerator and Memory connectivity scenarios for the NVMe & CXL enabled disaggregated datacenter
  • Identify and document Commonalities and Differences between Compute, Accelerator, Storage and Memory appliance connectivity requirements
  • Explore the Cost, Bandwidth, Latency, Density, Distance expectations with Electrical and Optical solutions to meet the PCIe NVMe & CXL requirements
  • Produce detailed external intra-rack PCIe (NVMe & CXL) connectivity High-Level scenarios and requirements

SSD

HDD

Memory

Seagate Internal

6 of 13

SFF Specification Charter 

  • Explore, study and specify optimal cable and connector solutions for Compute, Storage, Accelerator and Memory intra-rack connectivity scenarios and requirements described by the OCP workstream
  • Identify Commonalities and Differences between NVMe & CXL appliance connectivity requirements 
  • Investigate existing industry cost/performance optimized optical and copper cables & connectors to leverage volume production cost optimized components 
  • Produce detailed external PCIe (NVMe & CXL) detailed intra-rack connectivity specification 

SSD

HDD

Memory

Seagate Internal

7 of 13

NVMe & CXL PCIe Cabling Market Analysis

Identified two Intra-Rack markets for PCIe Cables w/ ~$600M/year potential revenue:

  1. NVMe DAS SSD & HDD Storage JBOD connectivity to Compute Nodes 
    • HDD Storage Market by 2025 ~3 Zettabytes
    • SSD Storage Market by 2025 ~1 Zettabyte
    • ~200Mu HDDs plus ~50Mu SSDs are sold Yearly
    • >200Mu will go in JBODs/JBOFs with ~50 Device per box 🡪 ~4Mu JBODs
    • Datacenters update their storage infrastructure once each ~4-5 years
    • Assuming ~4 cables per JBOD/JBOF we expect ~4Mu cables per year
      • AOC represent 1/4th of the cables over >3M length at ~$200 per cable 🡪 ~$200M
      • AEC represent 1/4th of the cables 1M-3M length at ~$100 per cable 🡪 ~$100M 
      • DAC represent 1/2 of the cables ~1M  length at ~$50 per cable 🡪 ~$100M
  2. CXL Pooled Memory connectivity to Compute Nodes
    • Market size is still unknow, but expected to require ~20 cables per memory pool
    • Assuming ~100K Memory Pooling Appliances per year 🡪 ~2Mu Cables/year
      • AOC represent 1/4th of the cables over >3M length at ~$200 per cable 🡪 ~$100M
      • AEC represent 1/4th of the cables 1M-3M length at ~$100 per cable 🡪 ~$50M 
      • DAC represent 1/2 of the cables ~1M  length at ~$50 per cable 🡪 ~$50M

Seagate Internal

8 of 13

Option-1: Changes Required to Current CDFP Solution

Direct Attached Storage(DAS) NVMe JBOD/JBOF as well as Compute Express Link(CXL) pooled memory expansion require external host node PCIe physical layer connectivity:

    • The older SFF8674 PCIe external connector not ideal beyond PCIe Gen4 x4
    • Changes are required to the CDFP connector with x8 and x16 to extend at PCIe Gen5 & Gen6 speeds beyond 1-Meter at 10(-12) BER in the form of:
      • Direct Attached Cable(DAC)
      • Active Electrical Cable(AEC)
      • Active Optical Cable(AOC)

Seagate Internal

9 of 13

Option-2: QSFP/QSFP-DD and/or OSFP/OSFP-XD

PCIe Gen5 Cabling:

  • x4 Full-Duplex Tx/Rx of 32Gbps using 56Gbps NRZ SR4 🡪 200Gbps QSFP
    • AEC & AOC Power budget of <2.5W per QSFP
    • Bit Error Rate(BER) of 10(-12) PCIe Gen5 Requirement
  • x8 Full-Duplex Tx/Rx of 32Gbps using 56Gbps NRZ SR8 🡪 400Gbps QSFP-DD/OSFP
    • AEC & AOC Power budget of <5W per QSFP
    • Bit Error Rate(BER) of 10(-12) PCIe Gen5 Requirement
  • x16 Full-Duplex Tx/Rx of 32Gbps using 56Gbps NRZ SR16 🡪 800Gbps OSFP-XD
    • AEC & AOC Power budget of <10W per QSFP
    • Bit Error Rate(BER) of 10(-12) PCIe Gen5 Requirement

PCIe Gen6 Cabling: 

  • x4 Full-Duplex Tx/Rx of 64Gbps using 112Gbps PAM4 SR4 🡪 400Gbps QSFP
    • AEC & AOC Power budget of <5W per QSFP-DD
    • Bit Error Rate(BER) of 10(-6) PCIe Gen6 Requirement
  • x8 Full-Duplex Tx/Rx of 64Gbps using 112Gbps PAM4 SR8 🡪 800Gbps QSFP-DD/OSFP
    • AEC & AOC Power budget of <10W per QSFP-DD
    • Bit Error Rate(BER) of 10(-6) PCIe Gen6 Requirement
  • x16 Full-Duplex Tx/Rx of 64Gbps using 112Gbps PAM4 SR16 🡪 1.6Tbps OSFP-XD
    • AEC & AOC Power budget of <20W per QSFP-DD
    • Bit Error Rate(BER) of 10(-6) PCIe Gen6 Requirement
  • Can’t use FEC for CXL, due to ns latency sensitive, but could use FEC for NVMe BER needs of 10(-12)

Seagate Internal

10 of 13

QSFP-DD Pinout Option

  • x8 Full-Duplex Tx/Rx
  • x4 Full-Duplex Tx/Rx (half-populated)
  • 4-pos Vendor Specific Definition (VSx)
  • Single-PCB solution
  • Support for DAC, AEC, and AOC
  • Support for 26-32AWG Twin-ax Cable
  • Available with Riding Heat Sink (RHS)

​

8.5mm

13.8mm

60mm

Form Factor Outline

Port� Profile

Seagate Internal

11 of 13

OSFP-XD Pinout Option

  • x16 Full-Duplex Tx/Rx
  • x8 Full-Duplex Tx/Rx (half-populated)
  • 4-pos Reserved For Future Use (RFFU)
  • Single-PCB solution
  • Support for DAC, AEC, and AOC
  • Support for 26-32AWG Twin-ax Cable
  • Available with Integrated Heat Sink (IHS) or Riding Heat Sink (RHS)

​

10.5mm*

22.8mm

95mm

*w/ RHS

Form Factor Outline

Port� Profile

Seagate Internal

12 of 13

Option-3: CW-WDM MSA Primarily for CXL

CXL 1.1/2.0:

  • Fully integrated electro-optical Chiplet supporting up to 8 WDM optical ports
    • Each port is a x8 link
    • Full-duplex Tx/Rx at 32 Gbps
    • Each port has 8λ per Tx/Rx fiber, based on the CW-WDM MSA
    • Port bandwidth is 256 Gbps, up to 2 Tbps per Chiplet
    • Bit Error Rate (BER) of 10(-12) meeting PCIe Gen5 requirement. No FEC required.
    • <1W per dual-port (x16) Chiplet
    • Re-driver-based Chiplet taking advantage of PCIe-Gen5 MR/LR SerDes on host ASIC
    • Latency around 1ns (electrical interface to optical interface)

CXL 3.0: 

  • Scalable architecture
    • Data rate at 64 Gbps
    • Support PAM4 modulation scheme
    • CW-WDM MSA supports 16λ and 32λ per fiber

Seagate Internal

13 of 13

Connect. Collaborate. Accelerate.

Seagate Internal