Highly efficient runtime scheduling and memory planning.
Multi-node and multi-GPU support. Scalable to 64 GPUs and more ….
Highly extensible modular design based on a novel notion of VM.