Modern-day high-performance computing (HPC) systems increasingly comprise disparate computing platforms (FFT processor, video processor, DSP processor, etc.) and a large number of general purpose cores. For many high throughput applications, HPCs need to run multiple applications in parallel, where each application is broken down to several tasks and multiple tasks are concurrently executed. However, given increasing heterogeneity of HPCs (many computing cores can run a task, but not all are always available) and complexity of workloads (increasing randomness due to memory access dependence, inter-task dependency, user and cloud-dependency, etc), scheduling of tasks becomes a challenging problem. Unlike the prior approaches where task scheduling is predetermined at the compilation time (therefore, rigid and not-adaptive to dynamically varying workload and computing resources) and done at software-level (therefore, slow), we present a hardware-driven approach where a dedicated accelerator proactively updates task schedules under varying resource availability and workload to maximize throughput of HPC..