ITME: Inference Tiered Memory Expansion with Disaggregated CXL-Hybrid Memories

This feed aggregates recent systems and ML-systems research (arXiv Jun 12 2026) focused on scaling and performance of large-model inference, multi-GPU/disaggregated memory architectures, device-cloud pipelines, and HPC optimizations. Highlights: ITME proposes using CXL-hybrid remote byte-addressable memory (validated with SK Hynix CMM, PCIe Gen5 NVMe SSDs and an FPGA prototype) to expand KV-cache capacity for LLM inference; Eidola extends gem5 to model fine-grained inter-GPU peer-to-peer traffic; a study shows limits of directive-based GPU portability (OpenACC vs OpenMP) across NVIDIA/AMD GPUs