In order to answer this question, we have to know something about SystemML's history, which began ten years ago in 2007 as a research project in the IBM Almaden Research Lab in California. The project was driven by the intention to improve the workflow of data scientists, especially those who want to improve and add functionality to existing machine learning algorithms.
So, SystemML is a declarative markup language that can transparently distribute work on Apache Spark. It supports Scale-up using multithreading and SIMD instructions on CPUs as well as GPUs and also Scale-out using a cluster, and of course, both together.
Finally, there is a cost-based optimizer in place to generate low-level execution plans taking statistics about the Dataset sizes into account. In other words, Apache SystemML is for machine learning, what Catalyst and Tungsten are for DataFrames.