Marvell’s Industry Analyst Day took place virtually on December 7, 2021. During the keynote, Marvell executives highlighted the datacenter networking portfolio with the latest Innovium and Inphi technology. Other products and announcements that stood out for us included:
4x400G MSA, announced on December 13th, 2021, offered an alternative path to 1.6T by starting with 16 of 100G lanes on both the optical and electrical sides of transceivers in the OSFP-XD form factor. Founding members of the new MSA, include Arista, Broadcom and Intel. We suspect that Amazon have encouraged these vendors to come up with a lower risk and faster path to market for 1.6T Ethernet transceivers. Modules based on the 4x400G approach will rely on proven optical and IC chips, currently used in 400G modules.
Figure below illustrates the three dimensions for charting a path to higher speed modules. $10 billion valuation of Inphi (now part of Marvell) was a step up along the vertical axis: from NRZ to PAM4. It started with 25G Baud rates for 200G and 2x200G transceivers and continued with 50G Baud rates for 400G (4x100G) and 800G (2x400G) transceivers. The next logical step is 100G Baud rate DSPs to support 8x200G design of 1.6T modules. This is a proven approach to reduce power consumption and cost by minimizing the number of components. Marvell has a clear path to 5nm and 3nm CMOS, 100G Baud rate DSP chips as well as financial resources to make it happen.
Several suppliers recently demonstrated the feasibility of 200G per lane optics, but the potential customers are concerned:
There is a good reason for the new MSA to be named 4x400G and not 16x100G, which would be another step along the “number of lanes” axis in the figure above, but that is not what 4x400G MSA is proposing. The idea is to take the existing 4-lane components and combine them into a high-density module. We have seen 2x200G and 2x400G double density modules already. This one will be the first quadruple density transceiver.
The new 4x400G MSA makes a step along the “hidden 4th dimension”, not shown in the figure. The distinction between 16x100G and 4x400G is very clear by looking at 4x400G FR4 modules: it is still a four-wavelength solution, but it needs 4 fibers in each direction. The differences are more subtle for DR4 devices, but the key point is that 4x400G modules will use 4x100G components, not 16x100G, ensuring a fast time to market.
If 4x400G modules make it to the market, we will have three generations of 1.6T transceivers:
Gen-2 will need 200G SerDes, which is another high-risk new technology, which will not be available until 2025. The sweet spot for Gen Zero is 2023-2024, Gen-1 2025-2026, Gen-2 2027 and beyond. The timeline is very tight, but if Amazon can deploy at least a million units in each of the three generations, the investment for suppliers is justified. More importantly, if Gen-1 and/or Gen-2 is delayed, there will be a manufacturable solution for 1.6T pluggable modules.The 4x400G transceivers may look like miniature porcupines with up to 32 fibers sticking out of them, but wait until we see co-packaged optics with 64 or 128 fibers. LightCounting will be releasing an updated forecast for CPO next week as part of its High Speed Cables, Embedded and Co-Packaged Optics Report - December 2021.