New CPU-GPU-AI Scheduling Approach Doubles Efficiency on Jetson Nano
• New Simultaneous and Heterogeneous Multithreading (SHMT) technique promises doubled performance and halved power consumption (4x efficiency) by utilizing CPU, GPU, and AI accelerator simultaneously • Proof-of-concept tested on Jetson Nano system showed 1.95x performance gains and 51% power reduction • Relies on quality-aware work-stealing scheduler to balance workloads and avoid errors • Software needs to be rewritten to take full advantage of SHMT parallelism • Performance gains dependent on problem size; smaller problems see less benefit