Topology optimization is a key computational tool in computational mechanics for generating high-performance, lightweight structures and is increasingly used in aerospace, automotive, and energy applications. Although supplying extensive design space, the curse of dimensionality restricts the widespread application of large-scale topology optimization in practical engineering. This work investigates the efficiency gains obtained by integrating high-performance computing with machine learning for ultra-large-scale topology optimization. We propose a parallel, problem-independent machine-learning (PIML)-enhanced framework. Built upon a substructuring (static condensation) strategy, once trained offline, the PIML model can be reused for different linear-elastic optimization problems with varying boundary conditions, mesh resolutions, and design domains. The proposed framework delivers three-level acceleration: (i) problem-independent machine learning model substantially reduces the effective dimension and computational cost of condensed stiffness matrix; (ii) high-performance computing (HPC) reduces the workload per process and enables large-scale parallel execution; and (iii) a parallel multigrid solver accelerates the coarse-grid equilibrium solve. Additional efficiency improvements are achieved via matrix-free condensation, direct condensation of uniform coarse elements, and practical tuning of computational resource limits. As a representative result, for a 3D topology optimization problem with 10.4 billion degrees of freedom (DoFs), the average wall-clock cost is just 42.0 s per iteration (8046 seconds in total) using 6,750 CPU cores, highlighting excellent scalability and computational performance. Weak scaling efficiency, strong scaling speedup, and maximum achievable efficiency are validated on multiple numerical examples, demonstrating a significant increase in tractable problem size and solution efficiency compared with conventional large-scale topology optimization workflows.