In 2016, G.fast looked very promising.
Thousands worked at developing and deploying.
It wasn't enough.
Most carriers are investing
in fiber or 5G instead.
Dark Blue: Firm commitments from incumbent: BT (10M), Belgacom, Australian NBN, Swisscom, Austria, Bezeq Israel, Chunghwa Taiwan, Telus Canada, Telekom South Africa, SK Korea, (U.S.) AT&T, Century, Frontier, Windstream, Belgium, Omantel
Mid Blue: Smaller carriers in Germany, Norway, Finland, Japan
- Published: 04 June 2016 04 June 2016
Brilliant engineers disagree. As g.fast extends to 212 MHz, noise on the line becomes increasingly difficult to tame. There is more noise than signal. Dong, of Huawei, had presented the idea of non-linear pre-coding previously. A respected engineer told me he was on target. May 20, his Paris presentation was convincing.
Jochen Maes of Nokia Alcatel came next to the podium and contended that non-linear coding was unnecessary. Maes works on the five gigabit Alcatel systems that go to 500MHz. He’s close to the data and believes what he calls “optimized linear” coding is actually best. He also believes that optimized linear requires far fewer calculations. The chips would require less power and run cooler, important factors.
Hard to characterize “bad” lines are a prime area of disagreement. Maes introduced some new tools to model the impact of channel perturbations. His results supported linear. Maes writes,
"All analysis on NLP so far assume theoretically ideal conditions. My presentation is the first that analyses the impact of channel fluctuations, e.g. due to mechanical vibrations or temperature fluctuations or imperfect channel estimates. It shows that NLP is more sensitive to such fluctuations and thus that under practical conditions NLP loses its advantage."
Because of the complexity, it is difficult to determine which factors are crucial. I am not an engineer. Even if I were it would be hard to decide who has this right. Definitive answers will not be available until there are extensive field trials. Even then, the question of the robustness of each implementation is likely to muddle the argument.
I remember the "QAM" versus "DMT" debate in VDSL a decade ago. That was an intense worldwide battle, with Texas Instruments and Alcatel employing a high-profile pr agency to protect their royalties. There were $hundreds of millions at stake. This one was bitter.
A prominent TI engineer accused me of being in the pay of the QAM team. That was surprising. I had written that it seemed to me DMT was technically superior. The only money I took from any of the parties was very visible advertising - from companies on the DMT side. (She never apologized.)
The QAM guys, Metalink and Infineon, had inexpensive chips ready to go, proven interoperability and support from China Telecom and others. They simply asked to be included as an option. Off the record, they contended the tests were unfair; the DMT folks had brought in John Cioffi at the last minute to tweak the system. (That may not have been true. I never asked John.) As I recall, after the QAM guys did their own tweaks, the result was very close.
I wouldn't be surprised if this debate goes on for a long time. Meanwhile, the chipmakers are in a chicken and egg situation. Until they incorporate new algorithms, it will be hard to test which coding offers better performance. Without a clear choice, it's hard for them to invest in adding the functionality to the chips.
I welcome on or off the record comments from all and have asked the principals for more details.