Inho Song | Searcharxiv

arXiv2026

Thermal Phase Transition of Three Dimensional Orthogonal Metals

Orthogonal metals provide a paradigmatic example of strongly entangled metallic states whose transport and thermodynamic properties resemble those of conventional Fermi liquids, while their single-electron excitations remain fundamentally distinct. Despite considerable interest in their unusual ground-state properties, orthogonal metals in three spatial dimensions and their finite-temperature behavior remain relatively unexplored. Here, we investigate a three-dimensional orthogonal metal and characterize the thermodynamic and spectroscopic signatures associated with its finite-temperature phase transition. Starting from a simple model of fermions coupled to spin degrees of freedom, we perturbatively demonstrate that the thermal transition known to occur in the purely bosonic limit persists in the presence of coupling to a gapless Fermi surface. We then examine physical observables that change prominently between the coherent and orthogonal-metal regimes. In particular, we calculate the evolution of the physical-electron spectral function and identify characteristic signatures in angle-resolved photoemission spectroscopy. We find a pronounced contrast in low-energy spectral weight between the coherent and orthogonal-metal regimes without requiring conventional symmetry-breaking order. Our results establish finite-temperature orthogonal-metal physics as a possible mechanism for pseudogap-like behavior in three-dimensional and quasi-two-dimensional layered materials with non-negligible interlayer hopping, which may be applied to the recent experiments in $\mathrm{Sr_2 VO_3 Fe As}$.

cond-mat.str-el↗

arXiv2026

A CXL Memory Rack for Multi-Turn LLM Serving

Long-context, multi-turn, and agentic LLM workloads increasingly reuse previously processed context, making KV-cache reuse essential for reducing redundant computation. However, this reuse shifts the bottleneck to the memory tier that stores and serves reusable KV states at cluster scale. GPU HBM and host DRAM are too costly to scale to TB-scale shared context capacity, motivating remote tiers built from lower-cost, higher-capacity media. This paper presents HyMCache, a CXL memory rack for multi-turn LLM serving. We build the memory rack using cost-efficient CXL-hybrid memory (CXL-HM), which combines a small amount of in-device DRAM with large SSD-backed capacity behind a CXL interface. By exploiting the read-dominant, predictable, and append-only nature of multi-turn KV-cache access, HyMCache rethinks DRAM management within CXL-HM to efficiently support TB-scale SSD-backed KV reuse. It uses request-level prefix prefetching and opportunistic write buffering to stage latency-critical reads in device DRAM, enabling DRAM-scale KV-cache efficiency at SSD-level cost. We evaluate HyMCache on a real CXL-HM prototype under both single-aggregator and PD-disaggregated serving configurations. Under the same DRAM budget, HyMCache outperforms local LMCache by 3.0x in single-node serving and 1.45x in PD-disaggregated serving. Compared with 1 TB distributed-DRAM Mooncake, HyMCache incurs about 30% lower performance but uses 16x less DRAM.

cs.DC↗