The Business Problem

A leading semiconductor company needed to optimize the design verification process for CPUs and GPUs. Traditional verification methods used random or manually-designed stimuli/payloads, which were inefficient at finding design bugs and corner cases.

The challenge was to develop a Machine Learning/Data Science workflow that could automatically generate optimal stimuli that thoroughly test CPU/GPU designs, uncovering bugs more efficiently than traditional methods.

The INM Consulting Approach

As project lead and principal developer, we developed a comprehensive ML/DS workflow using advanced generative models and reinforcement learning.

Implementation Details

• Implemented generative ML models (GANs) to generate diverse test stimuli
• Developed Reinforcement Learning (RL) models to generate optimal stimuli for CPU/GPU verification
• Machine Learning models developed using Keras/TensorFlow in Python
• Models monitored and versioned using MLflow
• Integrated work from 3 other developers into an end-to-end data product
• Data pipeline implemented in Linux, deployed on-premise
• Parallelized using Slurm for high-performance computing

Technologies Used