Micron Technology Clinic Effect of Transistor Number and Hierarchy on Simulation Speed Presented by: Jason Oneida Advisor: Dr. Ken Stevens
Overview ► Advantages of Two-Phase Analysis? Simulation speed Accurate abstraction Hierarchy in simulation ► Experiment: Cascading multipliers Test importance of hierarchy in design Scaling hierarchical mesh of 16-bit multipliers Implementation notes: Galois LFSR ► Data comparison HSimPlus and HSPICE Conclusions regarding Two-Phase Analysis
Experimental Setup ► Hypothesis: Hierarchy in execution will result in faster run-time Two-Phase Analysis and hierarchy Micron single-layer package model ► Metric for comparison Number of transistors vs. simulation time [1] “Central Processing Unit,” schools-wikipedia.org, Jan – Jan [Online]. Available: wikipedia.org/wp/c/Central_processing_unit.htm[Accessed: March 29, 2010] wikipedia.org/wp/c/Central_processing_unit.htm
Circuit Design ► Cascaded multipliers Hierarchy: divergent data path Non-ideal modeling ► Pseudo-random input Linear feedback shift register (LFSR) ► Galois LFSR [2] “Linear Feedback Shift Register,” absoluteastronomy.com, Jan, [Online]. Available: [Accessed: March 29, 2010]
Multiplier Block Diagram
HSimPlus Results ► Plot features Variables Scale ► Data Points Eight total ► Curve Fit Initial region Linear Portion
HSPICE Results ► Axes Units ► Simulation time 10,000 sec. per every 100,000 transistors ► Curve Shape Slope Fit trend
Combined Results ► Data point discrepancy – 8 vs. 6 ► Speed difference
Sample Multiplier Output ► HSimPlus settings ► Values: approximately 5% error
Conclusions ► Experiment to show benefit of hierarchy on run-time using two phase models ► Validated with a scalable design Hierarchical grids of multipliers Pseudo-random number generators for data ► Hypothesis confirmed HSimPlus scaled at about 1,000 sec. per 100,000 transistors HSPICE scaled at about 10,000 sec per 100,000 transistors Output accuracy within 5%