statistical-blockade

This repo contains an implementation of Statistical Blockade: A Novel Method for Very Fast Monte Carlo Simulation of Rare Circuit Events, and its Application.

Research paper can be found here

Problem

The delay distribution of a gate strongly depends on the device geometry and doping profile. In nano-scaled CMOS devices, the random variations of dopant atoms in the channel region cause random variations in the transistor threshold voltage (Vt), known as “random (or discrete) dopant effect”. This result in threshold voltage mismatch between transistors on die and significant delay variation of logic gates and circuit. Hence, a statistical modeling and analysis of the delay of logic gate is necessary. In this project, we consider random variation in transistors, and apply stochastic modeling to estimate distribution of propagation delay in given circuit.

Consider a critical path made with 5 inverter and 5 NAND cascaded as below. With variations in MOSFET, try to estimate the yield rate of circuits that have a delay time less than 1.395e-10s. In another word, try to estimate the probability of high sigma event that the time delay of the circuit is equal to or larger than 1.395e-10s.

The MOSFET used in circuit are shown as below. Both components should follow a design in BSIM3 model, which is given in the project src/ folder.

Prerequisites

• MATLAB
  • Statistics and Machine Learning Toolbox
• HSPICE

Generate Monte Carlo Training Data

1. In main.m, change varilable sample_training to the amount of samples you'd like to train on.

2. Run main.m

• Two output files will be generated training_fail_samples and training_pass_samples.

Train Classifier

*Need Statistics and Machine Learning Toolbox

1. Run classifier.m

• This will save the classifier tail predictions into saveTail.mat for use later.

Simulate Using HSPICE

1. Run sim.m

• This will use the classifier tail predictions from saveTail.mat.

• In sim.m:

  • Setting the threshold variable and you'll notice the change in the error count.
  • You can change the gppdf to gpcdf to get GDP CDF's plot.

Contributors

• Lini Mestar
• Yan Chen