Slide 1: From Reactive to Predictive: AI-Driven Optimization for ATE Performance & Reliability
From Reactive to Predictive: AI-Driven Optimization for ATE Performance & Reliability Wai-Kong Chen, Cohu, Inc. February 11, 2026 © 2026 Cohu, Inc. All rights reserved.
Slide 2: Agenda
Agenda • Reactive Automated (ATE) application - Pre-emphasis setting optimization • Predictive Automated (ATE) application - Improvement in ATE maintenance
Slide 3: Reactive: Automated Pre-Emphasis Setting
Reactive: Automated Pre-Emphasis Setting
Slide 4: Reactive: Automated Pre-Emphasis Setting
Reactive: Automated Pre-Emphasis Setting • Objectives - Provides a structured, automated method to select optimal pre-emphasis settings. - Combines simulation, measurement, and software-guided testing. - Supports a wide range of probe card and load board configurations. - Ensures robust performance
Slide 5: Pre-Emphasis Setting Challenges: Too Many Settings
Pre-Emphasis Setting Challenges Too Many Settings • There are 3 knobs for pre-emphasis setting - Pre-1 - Post-1 - Post-2 • Each knobs has 32 settings - A combination of 32K settings with 3 knobs - To determine the optimum setting will be time consuming and challenging without oscilloscope • Pre-emph
Slide 6: Hardware Characterization
Scope-Based Validation Hardware Characterization Signal Integrity Simulation Preliminary Setting Table Simulation Result Scope-Based Validation • To reduce the number of settings • Create Final Setting Table
Slide 7: Software-Based Optimization
PCB Trace Routing Software-Based Optimization • On the user load board, select an unused TX (transmit) lane that is routed similarly to other high-speed lanes. This ensures it accurately reflects the overall high-speed channel characteristics. • Create a loopback connection by routing the selected u
Slide 8: The Methodology: Real world data collection
The Methodology The base: Real world data collection The Scaling Rules • Half Rate Rule: If speed drops by 50% (6.25G), loss drops by ~50%. Steps scale by 0.5. • Voltage Ratio Rule: If swing drops by 50% (200mV), the required cancellation voltage drops by 50%. Steps scale by 0.5. Parameters & Relati
Slide 9: The "Sweet Spot" Chart
The “Sweet Spot” Chart • The map visualizes the Maximum Eye Quality achievable for every combination of Pre1 and Post 1. • By collapsing the other dimensions (Post2, RX settings), we can clearly see the “mountain peak” of signal integrity - Yellow/Green areas: High Eye Quality (the optimal zone) - P
Slide 10: Signal Integrity Landscape Visualization of Conditions
Signal Integrity Landscape Visualization of 5 Conditions • Based on the data points we have analyzed and the physical scale rules. The optimal emphasis setting (for both Pre1 and Post1) is directly proportional to the product of the Data Rate and Voltage Swing • The Wire mesh represents the “Eye Qua
Slide 11: Figuring Out the Formula
Figuring Out the Formula The data that generated the coefficient is accumulative For instance: Coefficient -0.0016 is calculated using dataset1. Coefficient -0.00187 is using dataset1 + dataset2
Slide 12: Pre-Emphasis Setting Formula
Pre-Emphasis Setting Formula X-Axis Value from the equation (e.g., -8.0) Y-Axis The specific setting for each tap Key Relationships Tracks the "Optimal Step" almost 1:1. If the Post-Tap 1 (Green) equation says -8, set Post1 to -8. Tracks closely but is slightly weaker. If the Pre-Tap (Blue) equation
Slide 13: Initial Results (based on actual data collected)
Initial Results (based on actual data collected) Data was manually collected using multiple pre-emphasis settings, data rates, and voltage swings Pre-emphasis settings were recommended by an AI model using the collected data. Validation results indicate strong correlation between the AI-suggested an
Slide 14: Predictive: Improving Cpk Value
Predictive: Improving Cpk Value
Slide 15: Predictive: Process Stability Monitoring
Predictive: Process Stability Monitoring Use Case Potential Impact • High-speed digital instrument: Low Cpk in • Difficulties in manufacturing voltage swing test • Out of box quality • Tester-to-tester or site-to-site variation • Insufficient design margin • Parameter sensitivity • Corner condition
Slide 16: Reason for Low Cpk (<2.0)
Reason for Low Cpk (<2.0) Process Mean is Shifted High Process Variation Test Condition Calibration Toward Upper Limit (Sigma) Sensitivity Margin The mean is closer to the This is the lowest swing volt The test The standard deviation is upper specification limit age test point (example of uses perce
Slide 17: Recommendations to Improve Cpk
Recommendations to Improve Cpk Center the Reduce Review Hardware Process Mean Process Variation Specification Limits Investigation • Adjust calibration offset to • Investigate sources of • The current limits are overly • Potential hardware-related shift the mean variation in the swing conservative f
Slide 18: Thank you
SEMICON’ & sem KOREA Thank you TRANSFORM A=Cohu: TOMORROW