De-correlating Training Patterns in Multi-Lane High-Speed Links
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High-speed Ethernet links with multiple lanes face challenges in de-correlating training pattern sequences, leading to increased noise and error rates due to cross-talk interference, as conventional random seed techniques are insufficient for higher-speed links like 100Gbps.
Innovation Solution
Employing different PRBS11 polynomials for each lane and dividing them into groups to ensure unique sequences at each endpoint, minimizing correlation and preventing false adaptation during link training, thereby reducing noise and error rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional random seed techniques are used for training patterns in high-speed multi-lane links, then link training can be performed, but cross-talk interference increases and error rates rise due to correlated sequences between lanes
Solution Approach 1:
The patent applies local quality by making each lane's training pattern unique through different PRBS11 polynomials and random seeds, so that each lane has distinct local characteristics rather than using identical patterns across all lanes. This differentiation reduces cross-correlation and minimizes cross-talk interference between adjacent lanes.
Solution Approach 2:
The patent changes the parameters of the training patterns by using different PRBS11 polynomials (e.g., x^11 + x^9 + x^5 + x^4 + 1 vs. x^11 + x^8 + x^7 + x^6 + 1) and different random seeds for each lane. This parameter variation ensures that training patterns remain uncorrelated across lanes, thereby reducing cross-talk effects while maintaining the required training functionality.
2Object-affected harmful factors
If different PRBS11 polynomials are used for each lane to reduce cross-talk, then cross-talk interference decreases, but the complexity of generating and managing unique sequences increases
Solution Approach 1:
The patent segments the training pattern generation by dividing the lanes into groups and assigning different PRBS11 polynomials to each group. This segmentation approach allows for systematic management of unique sequences while reducing the overall complexity compared to generating completely independent random patterns for each lane. The structured assignment of polynomials and seeds simplifies the generation process.
Data Source
Figure 1~9
Figure 1a~11
Figure 2a~2b
AI summary
Methods, apparatus and systems for de-correlating training pattern sequences for high-speed links and interconnects. The high-speed links and interconnects employs multiple lanes in each direction for transmitting and receiving data, and may be physically implemented via signal paths in an inter-plane board such as a backplane or mid-plane, or via a cable. During link training, a training pattern comprising a pseudo random bit sequence (PBRS) is sent over each lane. The PBRS for each lane is generated by a PBRS generator based on a PRBS polynomial that is unique to that lane. Since each lane employs a different PRBS polynomial, the training patterns between lanes are substantially de-correlated. Link negotiation may be performed between link endpoints to ensure that the PBRS polynomials used for all of the lanes in the high-speed link or interconnect are unique. Exemplary uses include Ethernet links, Infiniband links, and multi-lane serial interconnects.