Ethernet Port Speeds:
40G ports are in decline, while 100G ports are experiencing rapid growth. 100G is the fastest growing Ethernet standard in terms of revenue growth ever. 400G ports will start ramping in late 2018 and will become more popular in 2019 and 2020. 100G will continue to ship at high volume until around 2021, when 400G will likely take over.

Optics and Cabling:
Today, cloud networks use 100 GHz optics with copper cables in the rack, AOC cables from the top of rack to the leaf switch, and PSM4/CDM4 to the spine. For metro networks, LR4 or DCO optics are used. The 400G picture is similar, with cloud customers able to retain existing wiring or fiber installations and simply upgrade the speed per link by 4x. Copper cables and AOC cables will be used on the bottom, DR4/FR4 optics will be used to the spine, and DCO optics will be used in the metro.

MSA and Optics Landscape:
IEEE standards have traditionally taken a long time to finalize, leading to debates on optics specifications. The 16x25GB and 8x50GB optics initially proposed were not widely adopted.

100GB Lambda MSA:
A new Multi-Source Agreement (MSA) called 100GB Lambda MSA was formed to address the missing optics specifications in the IEEE standard. The MSA aims to define 100GB per lane optics for various distances, including 2km, 10km, and 40km.

Charter and Goals:
The 100GB Lambda MSA charter focuses on defining optics for 100GB per optical lane, with an emphasis on lower cost, lower power, and forward compatibility with future 100GB electrical lanes.

IEEE vs. MSA Optics:
The IEEE standard included 100GB DR and 400GB DR4 optics, while the MSA is defining the missing specifications.

Migration Path:
Customers can transition from current 100G CDM4 or PSM4 optics to 100G FR4 and DR4 using the same fiber infrastructure.

Lane Differences:
8-lane standards (LR8 and FR8) have different optical characteristics and are not interoperable with 4-lane standards, due to their use of 8 wavelengths instead of 4.

Future Ethernet Chip Lanes:
The next silicon generation will introduce 50 gigabits per second lanes, enabling 100 gigabits per second ports with only two lanes. The roadmap includes 100 gigabits per second electrical lanes expected in two years, allowing for 100 gigabits per second single-lane capabilities.

Market Trends:
10 gig lanes are declining, with 25 gig lanes becoming the bulk of shipments by next year. 50 gig PAM4 will peak in 2018 and start declining in 2019-2020. 100 gig kicks in late 2020, ramping to higher volumes in 2021-2022.

IEEE Standards and Electrical Lane Specs:
The IEEE takes three years to define new standards, with no current standard for below 400 gig electrical lanes. 50-lane definitions exist, and an eight-lane definition maps to 800 gig. The electrical lane spec, OIF-CIA-112G-VSR, defines the interface from switch silicon through PC board to the optical module.

Matching Silicon Lanes to Optical Lanes:
Silicon transitions drive optics transitions, avoiding slower speed optics tied to faster silicon. Inverse gearboxes for slower optics require more power and board space. Forward gearboxes are used when optics are faster than silicon speed or matching silicon speed.

Power Consumption and Cost Considerations:
25 gig silicon with C114 optics consumes ~4 watts (not counting silicon). 50 gig PAM4 chip with inverse gearbox increases power to ~20 watts per 100 gig channel. 100 gig DR with gearbox has lower power than CLM4. 2×50 and 100G cases have lower power consumption due to fewer lanes and simpler retimer.

Module Level Considerations:
Initial 400G 8×50 module with FR4/DO4 optics consumes ~10W at 12V. With 100G serial interface, power per channel drops to ~6W. Two channels can be combined into one module for cost-effectiveness.

OSFP Form Factor:
OSFP is an EMSA standard form factor for high-density pluggable modules. It can handle 5600-1200 or 1500GB PAM-4 lanes. OSFP can deliver up to 28 terabits in a single box with 100 gig interfaces, consuming up to 15 watts. It supports metro reach copper cables and digital core and optics configurations.

DCO Chips and Optics:
DCO chips in development will drive 400GB 16-QAM, consuming less than 15 watts. DR4, FR4, and LR4 optics are used for 8-fiber parallel and duplex fiber. Dual versions with two connectors will double the capacity for 800 gig applications.

Comparison of OSFP and QSFP-DD Form Factors:
OSFP and QSFP-DD are competing form factors for high-density pluggable modules. Both support 8 lanes and 36 ports per 1U, with backward compatibility to QSFPs using a mechanical adapter. Thermal and electrical performance differences exist, with QSFP-DD facing challenges at 50 GHz and potentially 100 GHz.

Thermal Capabilities:
QSFP and OSFP modules differ in thermal capabilities. OSFP offers a 30-40% better thermal capability compared to QSFP, critical when plugging in Digital Coherent Optics (DCOs).

Riding Heat Sinks and Thermal Interface:
QSFP modules rely on riding heat sinks, whereas OSFP modules have built-in heat sinks. The thermal interface between the QSFP module and the riding heat sink can be inconsistent, leading to a potential thermal gap of 50-60%.

QSFP’s Interior Volume Limitations:
QSFP modules lack sufficient interior volume for advanced optics. To address this, a Type II QSFP module was developed that extends 15 millimeters beyond the Type I, but it lacks heat sinks and controlled airflow outside the chassis.

Issues with QSFP PDD:
Andy Bechtolsheim views the QSFP PDD as a one-trick pony with a short lifecycle. Comparing it to the successful SFP, which underwent several upgrades from 1G to 10G, 25G, 50G, and 100G, he questions QSFP’s long-term viability.

QSRP’s Scaling Potential:
QSRP, initially a 10G interface, has evolved to support 40G and 100G, with the potential to scale to 200G and even 400G by 2020.

OSFP’s Superior Density:
OSFP offers double the density of QSFP and can support either 2x400GB or 800GB interfaces, while QSFP is limited to 8x50GB.

100GB DR and FR:
Next-generation silicon will transition to 50G PAM4, enabling support for the old CWM4 optics with an inverse gearbox. However, this approach consumes significant power (200 watts per 100G), prompting the adoption of forward gearboxes and lower-power optics. Honecke DR and FR optics are expected to deliver significantly lower power and cost savings.

Power Savings with 2×50:
Using 2×50 optics, along with a forward gearbox, can significantly reduce power consumption compared to the 4×25 approach, offering substantial energy savings.

AOC Cables and Fiber Requirements:
The transition to 2×50 requires new AOC cables, copper cables, duplex and multimode fiber, and duplex single-mode fiber to replace the current optical setup on Honeygeek.

Important Developments in Optics for High-Speed Networking:
Two new form factors are being proposed for high-speed networking: MicroQSFP and SFPDD. MicroQSFP has four lanes, but can use just two, providing good electrical and thermal performance. SFPDD resembles an SFP but adds another row of connectors for additional channels, but has weaker electrical performance and thermal challenges. There may be two transitions in the market: from 4×25 QSFP to 2×50 form factors, and then from 2×50 to 1×100 SFP (typo corrected from “by 100”).

Market Trends and Challenges:
The volume of 100GB DR optics is projected to exceed the historical volume of 4LAN by 4×25. Optics vendors are struggling to meet the demand for 100GB optics and need to find efficient ways to produce these modules. DR1 modules are easier to build than 4LAN modules in terms of optical assembly and yield. Use cases for 100GB optics include top-of-rack connections, where copper cables become challenging.

Economics and Adoption:
Lower cost and lower power consumption drive adoption of new technologies in networking. DR1 and DR4 optics offer a cost-effective way to address the market demand. DR4 optics can be split into four channels of DR1 optics using a splitter cable.

Transition from NRZ to PAM-4 and 100 Gigabit Electrical Lines:
The migration from NRZ to PAM-4 was a difficult and lengthy process, taking over 18 months after the peak of 40-gigabit adoption. The transition to 100-gigabit electrical lines is expected to happen within two years.

Customer Behavior and New Product Adoption:
Customers quickly adopt products that offer better performance and cost savings. The transition from 10 gig to 25 gig and 40 gig to 100 gig occurred quickly due to the advantages of the newer products. When a new product is better in every metric and the software works, customers adopt it rapidly. The transition to 50 gig will follow similar patterns, with optics needing to follow suit to fully realize the benefits.

Electrical Trace Limitations:
Current electrical traces on PC boards have limitations in terms of signal loss and reach. Higher speed signals experience more significant losses, limiting the achievable data rates. This constraint makes it challenging to build systems with electrical interfaces beyond 100 Gbps.

200-Gig PEM-4 Standard:
Some discussions have centered around a 200-Gig PEM-4 standard. However, practical implementation reveals that such a standard would be limited to very short distances due to signal loss within the packaging.

Multi-Chip Modules:
Multi-chip modules with embedded optics are being considered for future development in 2021-22. This approach could potentially address the limitations of electrical interconnects.

Transition to Photonics:
The limitations of electrical interconnects are driving a trend towards photonics for optical interconnections. Challenges with laser reliability and temperature sensitivity need to be addressed for widespread adoption of photonics.

Modulators on Chip:
An alternative approach is to place modulators on the chip while keeping lasers external and making them redundant. This technical solution requires new packaging at the chip level and extensive testing before it becomes viable.

Arista’s Approach:
Arista relies on third-party DSPs, such as Acacia, for its current products. The company does not engage in chip development but instead leverages the investments made by the broader industry. Arista selects and integrates the first available silicon that meets customer requirements.