Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2504.14893

Hardware-based Heterogeneous Memory Management for Large Language Model Inference

Soojin Hwang·Jungwoo Kim·Sanghyeon Lee·Hongbeen Kim·Jaehyuk Huh
View original source ↗Read source document ↗Export metadata ↓

Abstract

A large language model (LLM) is one of the most important emerging machine learning applications nowadays. However, due to its huge model size and runtime increase of the memory footprint, LLM inferences suffer from the lack of memory capacity in conventional systems consisting of multiple GPUs with a modest amount of high bandwidth memory. Moreover, since LLM contains many bandwidthintensive kernels, only focusing on the memory capacity without considering the bandwidth incurs a serious performance degradation. To handle such conflicting memory capacity and bandwidth demands in a cost-effective way, this study investigates the potential of heterogeneous memory systems, proposing H2M2. It uses an asymmetric memory architecture consisting of capacity-centric and bandwidthcentric memory with computation units attached to each memory device. With the asymmetric memory, we first analyze the effect of kernel-memory mapping for the asymmetric memory. Second, we propose a dynamic runtime algorithm that finds a mapping solution considering the characteristics of LLM operations and the change of footprint during LLM inference. Third, we advocate the need for memory abstraction for the efficient management of the asymmetric memory. H2M2 outperforms the conventional homogeneous memory system with LPDDR by 1.46x, 1.55x, and 2.94x speedup in GPT3-175B, Chinchilla-70B, and Llama2-70B, respectively.

Explore related subjects

cs.AR

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Soojin Hwang, Jungwoo Kim, Sanghyeon Lee, Hongbeen Kim, Jaehyuk Huh. 2025-04-21. Hardware-based Heterogeneous Memory Management for Large Language Model Inference. https://arxiv.org/abs/2504.14893

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

HLSmith: An Expert-Guided Agentic Framework for C/C++-to-HLS Translation

Application-specific FPGA accelerators offer substantial performance and energy-efficiency gains across many application domains, but developing them is costly, often requiring months of specialized effort. Even with high-level synthesis (HLS), designers still need extensive hardware expertise to build high-performance accelerators. Although large language models (LLMs) have demonstrated strong software-generation capabilities, even frontier models lack the hardware intuition and procedural knowledge needed to reliably translate baseline C/C++ programs into high-performance HLS designs: they struggle to identify effective architectures, follow the optimization processes used by HLS experts, and apply hardware transformations consistently across diverse kernels. We present HLSmith, an expert-guided framework for translating C/C++ programs into optimized HLS accelerators. HLSmith combines three components: an HLS optimization expertise library that encodes guarded transformation recipes, their applicability and prerequisite conditions, and unsafe cases to avoid; a staged, feedback-driven orchestration flow modeled on expert HLS development practice that guides agents through synthesis, bottleneck analysis, and optimization; and a tool-grounded model-adaptation pipeline that converts optimization trajectories from commercial frontier models into training data for fine-tuning open-weight LLMs. We evaluate HLSmith on PolyBench against ChatHLS, a leading prior agent-orchestration framework for HLS accelerator development. HLSmith achieves a geometric mean speedup of 4.24x over ChatHLS while producing functionally correct designs, in both software and RTL simulation, for every benchmark, compared with ChatHLS's 57% valid-design rate. It further reaches speedups of up to 252x and 138x with commercial frontier models and open-weight models, respectively.

cs.AR↗
arXiv2026

TEMPO: A Tag-Based Framework for Efficient Memory Ordering

Weak-memory processors rely on ordering instructions for correctness, yet conventional implementations often enforce them more conservatively than the memory model requires. This over-enforcement manifests as drain-induced retirement stalls at ordering instructions and conservative squash/replay of speculative loads, suppressing legal executions and reducing throughput. We present TEMPO, a tag-based framework for precise micro-architectural implementation of ordering instructions. TEMPO assigns lightweight ordering tags to instructions and decomposes enforcement across retirement-time predicates and completion-time store ordering, allowing the core to enforce required ordering without conservative retirement serialization. TEMPO eliminates unnecessary retirement serialization at ordering instructions and speculative-load squash/replay. In our evaluation, TEMPO reduces geometric-mean normalized execution cycles by 7.9% on native four-thread workloads and improves geometric-mean IPC by 15.9% on an instrumented SPEC2017 dynamic binary translation (DBT) proxy for cross- ISA execution (e.g., x86-on-Arm), while adding only 262 bytes per core.

cs.AR↗
arXiv2026

Accelerating the Mitigation of LLM Inference Nondeterminism Across GPU Architectures

Large language model (LLM) outputs are expected to be reproducible under greedy decoding, yet in practice the same model, prompt, and software stack produce different outputs on different GPUs. The root cause is floating-point non-associativity combined with hardware-dependent kernel selection. Inference frameworks select different matrix-multiplication kernels on each architecture, with different parallel reduction orders and unspecified tensor-core arithmetic, and the resulting rounding differences can flip output tokens. Existing solutions have imperfect cross-architecture reproducibility and incur a significant performance penalty. We present a solution employing a set of fixed-configuration fused-upcast GEMM kernels that load 16-bit weights from memory, upcast them to FP32 in registers, and accumulate with IEEE-754 arithmetic in a reduction order that is a pure function of the problem shape and is therefore independent of the device, its SM count, or kernel scheduling. By fixing the floating-point reduction order as a function of problem shape alone, every GPU runs the same operation sequence, so cross-architecture reproducibility of the linear layers reduces to correct IEEE-754 arithmetic rather than to rounding differences staying below a tie-flip threshold. We confirm our solution's linear-layer outputs are bitwise identical across NVIDIA Ampere, Ada, and Hopper GPUs, while running $1.17$ to $3.1\times$ faster end-to-end than the state-of-the-art solution and cutting weight-memory traffic in half.

cs.AR↗