Hardware-accelerated Aggregation: Unification and Specialization

Collection of new arXiv CS papers (announced 2026-06-10) covering systems and theory for ML and distributed systems. Highlights: a unified framework and hardware-specific optimizations for hardware-accelerated aggregation on CPU/GPU/FPGA; RATrain, a resource-aware runtime for dense LLM training on bandwidth- and memory-constrained heterogeneous supercomputers (MT-3000); ASTRA-sim 3.0 for high-fidelity distributed ML simulation with cache-line granularity and InfraGraph infrastructure representation; a CPU–GPU hybrid system to achieve cloud-grade SLOs for local Mixture-of-Experts (MoE) models (