[History and development of optical Ethernet] Examining the Electrical interface at 200G, halving the power consumption required for communication [Net new technology] --INTERNET Watch

2022-05-27 21:51:15 By : Ms. Joyce Wang

Regarding Ethernet or 10GBASE-T, from 2017 [10GBASE-T, finally popular?], We delivered a total of 11 times and 2 extra editions.However, Ethernet for copper wiring by twisted pair is up to 10GBASE-T, and 25 / 640GBASE-T has not yet been put into practical use.In the first place, there is no such module standard, and even if it is developed, it is essential that it be as large as CFP2 / CFP8, and if it is not good, it may be as large as CFP as shown below, and of course it is the current QSFP / OSFP. Compatibility with system modules will also be lost.It would be difficult to expand the signal terminals of the QSFP / OSPF system to 16 lanes in each direction.Returning to the story, that's why we also considered Electrical Signaling of 200G per lane.The presentation "Considerations on 200G per lane PAM Signaling" has been presented by Huawei's Yuchun Lu and Yan Zhuang.By the way, the story explained at the beginning is based on the module interface (AUI) in mind, but in reality, the connection between dies in the LSI inside the module and the wiring inside the module are also in view.On the contrary, Backplane / Copper Interconnect systems such as KR and CR are out of consideration here (although I am afraid that KR / CR is completely included in Power Efficiency).Probably, regarding KR / CR, the distance is too long with the same standard and it is difficult to reach it as it is, so it seems that it can not be realized without adding additional FEC etc., but in the first place it is possible to realize a standard with wiring less than 1 foot. It is obvious that the standard of 1m order will not be established without it, and as the basis of the KR / CR standard, it seems that we want to first formulate a wiring standard of less than 1ft or cm order.Here, it is mentioned that 200G is passed, more efficient (2 to 3pJ / bit) transmission, and compatibility between 100G and earlier standards are maintained.This compatibility will be considered first with FEC using RS (544,514) (which would be standard), but if necessary, more powerful FEC will be considered.With RS (544,514), the latency is around 100ns (200G will not reach 50ns), but with a more powerful FEC, the latency will increase further.Ethernet is a standard that emphasizes throughput over latency, so it may not be a problem at this level (at least much less latency than 10GBASE-T).How can this be achieved?It is a story that has already come out that it is up to about 25G if it remains NRZ.After that, it is necessary to use PAM4 etc., but if it is left as it is, SNR (Signal-to-Noise Ratio) can be compensated by RS (544,514) FEC.However, if PAM16 is used, SNR deteriorates extremely, and it is necessary to use extremely powerful FEC, which is not reasonable.The data rate can be doubled by setting the current signal, which is premised on Differential (differential signal), to Single Ended and then using PAM4 modulation for each.The method of canceling crosstalk between signals at this time is already widely used as a MIMO algorithm.And, in order to realize 200G per lane, this SE (Single Ended) PAM4 is better.Thunderbolt 5 is a standard that passes 80 Gbps with PAM3, and with PAM3 it can pass 3 bits with 2 transfers (since 3 values ​​of -1/0/1 can be taken per transfer, a total of 9 types can be taken with 2 transfers. It will be a combination, but since both cases of 0 are not treated as data, it will be 3 bits).In other words, the signal speed is 53.3GHz, and it is unclear whether FEC etc. will actually be included, but even if it is assumed that RS (544,514) is used as FEC like Ethernet and the overhead is added, the signal speed will be. It is a calculation that will be about 56.4 GHz.Technological advances are remarkable (although 200G lanes are still tough) as this speed has become a reality with copper wiring.Below is a comparison of PR PAM4, which combines PAM4 with a technique called Partial Response, and SE PAM4, PAM6, DSQ-32, PAM8, DSQ-128, and PAM16 (8 methods including the original PAM4).Looking at the DSQ-128 and PAM16, it seems that this Eye Height is extremely strict or impossible.To put it the other way around, I feel that 10GBASE-T often adopted this kind of thing.Looking at SE PAM4 like this, it seems that Eye Height is not so high, but this is because the two signals are arranged one above the other, and in fact, the same Height as PAM4 is secured.And, the examination of Signaling Power Spectrum Density is as follows, and it is said that PR-PAM4 / SE-PAM4 and PAM16 can reduce power consumption by half (at the same bit rate) compared to plain PAM-4. The result is that it serves the purpose of halving the power consumption required for communication.However, this is purely the power consumption of the communication path, and both PR PAM4 and SE PAM4 have more modulators than PAM4 (PR PAM4 adds Encode / Decode of the PR stage. , SE PAM4 requires two PAM4 modulators for one lane, that is, one-to-two signals), so the total power consumption cannot always be halved.Although this area needs a little more consideration, it is positive as a technical possibility.Speaking of SNR, based on PAM4, PR PAM4 / SE PAM4 is around 3dB, PAM6 / DSQ32 is around 3.5dB, PAM8 is around 6.3dB, DSQ128 is around 9.5dB, PAM16 is around 12.4dB. There is a penalty before and after.The lower the SNR penalty, the higher the feasibility, and in that sense, PR PAM4 / SE PAM4 is influential.When comparing Symbol Rate, Unit Interval, Nyquist Requirements, Bandwidth Requirements, and the penalties for ensuring the required BER, the result is SE PAM 4 with no constraints at all. ing.In response to this, it seems that SE PAM4 or PR PAM4 can realize an electrical interface, and moreover, it can be used not only inside modules and chips, but also for 200G KR / CR. There is.In my personal opinion, SE PAM4 is extremely vulnerable to external noise (in short, it's just a Single Ended) even if it can deal with crosstalk, so it may be strict for KR / CR. I feel like it.However, such a study will be conducted in the future.Anyway, this is just a guideline.Free technical writer.We have a wide range of fields of expertise, from CPUs, memories, and chipsets to communications, operating systems, databases, and medical care.The homepage is http://www.yusuke-ohara.com/OTN support for Beyond 400 Gb / s Ethernet was rejected by vote in AprilTrends in OSFP MSA and two IEEEs related to 800G EthernetInterconnect and transceiver standards defined by IEEE 802.3ba for 40G / 100GCopyright © 2018 Impress Corporation. All rights reserved.