I think you are training the network on gpu, instead try changing the "executionEnivronment" to "cpu" in Train Model section of the Aerial Lidar Semantic Segmentation Using PointNet++ Deep Learning example and try running it (for R2021a version).
In case you are now trying to run the example in R2021b, then in the Specify Training Options section, add "ExecutionEnvironment" Name-Value argument to the trainingOptions and set it to "cpu".
