Eye Diagram Computation of High Speed Links Using

Eye Diagram Computation of High Speed Links Using Machine Learning Majid Ahadi, Mourad Larbi, Madhavan Swaminathan Problem Statement and Challenges Example and Preliminary Results • The proposed approach uses Machine Learning (ML) techniques to find the eye diagram of high speed links with up to 100 X speedup. • Eye diagram shows performance of highspeed channels. • Bit error rate of high speed channels is very low. • Million of bits and a long time domain simulation is needed. • Example: • • • chip Wiring length = 20 Optimized receiver gain and peaking Baud rate = 16, FFE used, No DFE 1 e 7 and 1 e 6 bits on HSSCDR 1 e 6 bits for the ML approach. chip Hybrid-LGA Connector Inner Layer Wiring Processor Pkg PCB Vias • In the ML approach each sample is replaced with a kernel. • Kernel’s type is currently fixed to Gaussian. • Kernel’s bandwidth is determined by standard deviation of all samples. • Previous works suggested to reduce the cost: • • Statistical eye: Based on pulse response and superposition. Quick eye: Based on edge responses and superposition. IBM’s HSSCDR Stat Mode: Using S-parameters and convolution. These approaches have limited use and accuracy. They only work on linear systems and currently many developers use the traditional eye diagram. • Mentors on HSSCDR from IBM: Jose Hejase and Dale Becker. Proposed Approach • We believe further improvement in this method will reduce the error. • This approach directly finds the Bit Error Rate (BER) at receiver. • By learning the receiver voltage’s PDF, using an ML kernel estimation approach. PDF of Receiver voltage at a sample point Link to Uncertainty Quantification • Receiver voltage PDF: • • Part a: low stays low Part b: low turns high Part c: high turns low Part d: high stays high • Error happens when • a or b falls beyond zero • c or d falls below zero a b c d We Need Data! • Challenges in the kernel estimation approach PDFab Error • Kernel selection • Bandwidth selection • Proposed solution • Suggests kernel and bandwidth for similar problems, through ML classification. • Needs more data