~ similar to 2606.10440· 20 results
This paper introduces Eidola, a scalable extension to the gem5 simulation framework for modeling inter-GPU communication traffic.
The paper presents a field study on deploying large inference workloads on a non-GPU AI accelerator and identifies eight categories of limitations.
Yuncheng Yao, Zhenzhou Qi, Junyao Zheng, Chung-Hsuan Tung +2 more
This paper presents Beaver, a GPU sharing system that jointly manages compute and memory resources to protect the vRAN's strict processing deadline while serving AI workloads.
This paper presents Isaac-Net, a GPU-batched 5G New Radio module that simulates thousands of environments in lockstep with Isaac Lab physics.
Hawkeye is a system that allows perfect, precision-preserving reproduction of GPU-level matrix multiplication operations on a CPU, enabling efficient and trustworthy third-party auditing of machine le…
Yuchen Fan, Minghong Sun, Jikui Ma, Yunpeng Xu +18 more
The paper presents CHASE, an application-driven framework that explores physically feasible Cross-layer Heterogeneous System architectures for executing workloads with diverse requirements.
Megan Frisella, Shubham Tiwari, Andy Ruan, Yi Pan +4 more
Piper is a user-controllable distributed training system that decouples the strategy from the runtime implementation, allowing users to declare a comprehensive distributed training strategy with a sma…
Megan Frisella, Shubham Tiwari, Andy Ruan, Yi Pan +4 more
Piper is a user-controllable distributed training system that decouples the strategy from the runtime implementation, allowing users to declare a comprehensive distributed training strategy with a sma…
The paper proposes a fine-grained approach to overlap expert compute with the second all-to-all communication in Mixture-of-Experts (MoE) models, improving distributed MoE execution efficiency on mult…
RankLens is a single-rank emulation system that enables analysis, debugging, and tuning of distributed training with reduced resource requirements.
This paper presents MemX, a memory-aware automatic parallelism system for large-scale model training on cross-vendor heterogeneous GPU clusters.
This paper compares the effectiveness of moving inference state across sites versus recomputing it, and quantifies the sensitivity of this decision to various factors.
Daifeng Li, Huiqiang Jiang, Chengruidong Zhang, Wei Wu +5 more
The paper introduces D2K-Bench, a diagnostic benchmark to measure how effectively large language model agents translate expert design guidance into efficient GPU kernels.
Prasanna Date, Kevin Zhu, Shruti Kulkarni, Ashish Gautam +11 more
This paper introduces SuperNeuroMAT, a fast and scalable Python-based spiking neural network simulator that outperforms existing simulators and supports both dense and sparse modes.
Haishan Zhu, Domi Yan, Michael Levesque-Dion, Changxu Zhang +36 more
The paper presents the first production-scale application of Triton on a custom ML accelerator, MTIA-2i, and develops a compiler backend and language extensions for it.
The paper introduces Rotary GPU, an exploratory execution approach demonstrating that large Mixture-of-Experts models can be run locally on consumer GPUs with limited VRAM, achieving usable decode thr…
The paper introduces FrontGI, a set of techniques for fast and scalable GNN inference on large-scale graphs.
Taurus is a single-machine system for efficient Graph Neural Network (GNN) inference on large-scale graphs that do not fit in RAM, using source-centric broadcasts and a pipelined GPU-CPU-SSD hierarchy…