Efficient systems have made notable contributions to the recent success of machine learning. My research work has been mainly dedicated to improving the efficiency of machine learning, particularly with GPUs. In this seminar, I will present our work on accelerating Support Vector Machines (SVMs) and Gradient Boosting Decision Trees (GBDTs). The series of research work has led to open source projects: (i) ThunderSVM which is ~100 times faster and (ii) ThunderGBM which is ~10 times faster and more scalable to high dimensional problems than their counterparts. Faster systems may bring breakthroughs. Our recent results on a popular sentiment analysis problem show that our SVM based solution can achieve competitive predictive accuracy to the Deep Neural Network (and can even outperform the majority of the BERT) based approaches. Furthermore, our solution is about 40 times faster in inference and has 100 times fewer parameters than the models using BERT.