Serving Large Language Models on Huawei CloudMatrix384

• URL: http://arxiv.org/abs/2506.12708v3
• Date: Thu, 19 Jun 2025 12:27:10 GMT
• Title: Serving Large Language Models on Huawei CloudMatrix384
• Authors: Pengfei Zuo, Huimin Lin, Junbo Deng, Nan Zou, Xingkun Yang, Yingyu Diao, Weifeng Gao, Ke Xu, Zhangyu Chen, Shirui Lu, Zhao Qiu, Peiyang Li, Xianyu Chang, Zhengzhong Yu, Fangzheng Miao, Jia Zheng, Ying Li, Yuan Feng, Bei Wang, Zaijian Zong, Mosong Zhou, Wenli Zhou, Houjiang Chen, Xingyu Liao, Yipeng Li, Wenxiao Zhang, Ping Zhu, Yinggang Wang, Chuanjie Xiao, Depeng Liang, Dong Cao, Juncheng Liu, Yongqiang Yang, Xiaolong Bai, Yi Li, Huaguo Xie, Huatao Wu, Zhibin Yu, Lv Chen, Hu Liu, Yujun Ding, Haipei Zhu, Jing Xia, Yi Xiong, Zhou Yu, Heng Liao,
• Abstract summary: Traditional AI clusters face limitations in compute intensity, memory bandwidth, inter-chip communication, and latency.<n>This paper introduces Huawei CloudMatrix, a next-generation AI architecture realized in the production-grade CloudMatrix384 supernode.<n>It integrates 384 Ascend 910 NPUs and 192 Kunpeng CPUs interconnected via an ultra-high-bandwidth Unified Bus (UB) network, enabling direct all-to-all communication and dynamic pooling of resources.
• Score: 28.88558053380112
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: The rapid evolution of large language models (LLMs), driven by growing parameter scales, adoption of mixture-of-experts (MoE) architectures, and expanding context lengths, imposes unprecedented demands on AI infrastructure. Traditional AI clusters face limitations in compute intensity, memory bandwidth, inter-chip communication, and latency, compounded by variable workloads and strict service-level objectives. Addressing these issues requires fundamentally redesigned hardware-software integration. This paper introduces Huawei CloudMatrix, a next-generation AI datacenter architecture, realized in the production-grade CloudMatrix384 supernode. It integrates 384 Ascend 910 NPUs and 192 Kunpeng CPUs interconnected via an ultra-high-bandwidth Unified Bus (UB) network, enabling direct all-to-all communication and dynamic pooling of resources. These features optimize performance for communication-intensive operations, such as large-scale MoE expert parallelism and distributed key-value cache access. To fully leverage CloudMatrix384, we propose CloudMatrix-Infer, an advanced LLM serving solution incorporating three core innovations: a peer-to-peer serving architecture that independently scales prefill, decode, and caching; a large-scale expert parallelism strategy supporting EP320 via efficient UB-based token dispatch; and hardware-aware optimizations including specialized operators, microbatch-based pipelining, and INT8 quantization. Evaluation with the DeepSeek-R1 model shows CloudMatrix-Infer achieves state-of-the-art efficiency: prefill throughput of 6,688 tokens/s per NPU and decode throughput of 1,943 tokens/s per NPU (<50 ms TPOT). It effectively balances throughput and latency, sustaining 538 tokens/s per NPU even under stringent 15 ms latency constraints, while INT8 quantization maintains model accuracy across benchmarks.

Related papers

• AutoNeural: Co-Designing Vision-Language Models for NPU Inference [24.05617280495125]
  AutoNeural is an NPU-native VLM architecture co-designed for integer-only inference.<n>We replace the standard ViT encoder with a MobileNetV5-style backbone utilizing depthwise separable convolutions.<n>Our approach delivers substantial efficiency gains, reducing quantization error of vision encoder by up to 7x and end-to-end latency by 14x compared to conventional baselines.
  arXiv  Detail & Related papers  (2025-12-02T16:45:25Z)
• Eliminating Multi-GPU Performance Taxes: A Systems Approach to Efficient Distributed LLMs [61.953548065938385]
  We introduce the ''Three Taxes'' (Bulk Synchronous, Inter- Kernel Data Locality, and Kernel Launch Overhead) as an analytical framework.<n>We propose moving beyond the rigid BSP model to address key inefficiencies in distributed GPU execution.<n>We observe a 10-20% speedup in end-to-end latency over BSP-based approaches.
  arXiv  Detail & Related papers  (2025-11-04T01:15:44Z)
• MOBIUS: Big-to-Mobile Universal Instance Segmentation via Multi-modal Bottleneck Fusion and Calibrated Decoder Pruning [91.90342432541138]
  Scaling up model size and training data has advanced foundation models for instance-level perception.<n>High computational cost limits adoption on resource-constrained platforms.<n>We introduce a new benchmark for efficient segmentation on both high-performance computing platforms and mobile devices.
  arXiv  Detail & Related papers  (2025-10-16T18:00:00Z)
• CollaPipe: Adaptive Segment-Optimized Pipeline Parallelism for Collaborative LLM Training in Heterogeneous Edge Networks [57.95170323315603]
  We introduce CollaPipe, a distributed learning framework that integrates collaborative pipeline parallelism with federated aggregation to support self-evolving networks.<n>In CollaPipe, the encoder part is adaptively partitioned into variable-sized segments and deployed across mobile devices for pipeline-parallel training, while the decoder is deployed on edge servers to handle generative tasks.<n>To enhance training efficiency, we formulate a joint optimization problem that adaptively allocates model segments, micro-batches, bandwidth, and transmission power.
  arXiv  Detail & Related papers  (2025-09-24T07:54:01Z)
• eIQ Neutron: Redefining Edge-AI Inference with Integrated NPU and Compiler Innovations [4.776283807742058]
  eIQ Neutron efficient-NPU is integrated into a commercial flagship MPU.<n>Our solution achieves an average speedup of 1.8x (4x peak) at equal TOPS and memory resources across standard AI-benchmarks.
  arXiv  Detail & Related papers  (2025-09-17T19:45:51Z)
• Omniwise: Predicting GPU Kernels Performance with LLMs [0.06666419797034795]
  We introduce Omniwise, the first end-to-end, self-supervised fine-tuning pipeline that applies large language models (LLMs) to GPU kernel performance prediction.<n>It can predict key performance metrics, including memory bandwidth, cache hit rates, GFLOPs, and arithmetic intensity, directly from kernel code without the need for code execution or profiling tools.<n>Our approach achieves over 90% of predictions within 10% relative error on GPU kernels executed on AMD MI250 and MI300X architectures.
  arXiv  Detail & Related papers  (2025-06-25T23:36:44Z)
• FFN Fusion: Rethinking Sequential Computation in Large Language Models [16.8637819797503]
  We introduce FFN Fusion, an architectural optimization technique that reduces sequential computation in large language models.<n>We develop a principled methodology for identifying and fusing such sequences, transforming them into parallel operations.<n>Applying these techniques to Llama-3.1-405B-Instruct, we create an efficient and soon-to-be publicly available model that achieves a 1.71X speedup in inference latency and 35X lower per-token cost.
  arXiv  Detail & Related papers  (2025-03-24T17:20:35Z)
• A Novel Hat-Shaped Device-Cloud Collaborative Inference Framework for Large Language Models [12.644230479753476]
  Traditional cloud-based large language models (LLMs) meet high-accuracy requirements, but fall short of critical demands for low delay and enhanced privacy.<n>We propose HAT, a novel device-cloud collaborative inference framework that leverages the complementary strengths of U-shaped inference and speculative decoding.<n>We show that HAT achieves promising performance improvements, reducing TTFT by 41% to 54% and TBT by 41% to 77% compared to the baselines.
  arXiv  Detail & Related papers  (2025-03-23T10:54:58Z)
• EPS-MoE: Expert Pipeline Scheduler for Cost-Efficient MoE Inference [49.94169109038806]
  This paper introduces EPS-MoE, a novel expert pipeline scheduler for MoE that surpasses the existing parallelism schemes.<n>Our results demonstrate at most 52.4% improvement in prefill throughput compared to existing parallel inference methods.
  arXiv  Detail & Related papers  (2024-10-16T05:17:49Z)
• Hyperdimensional Computing Empowered Federated Foundation Model over Wireless Networks for Metaverse [56.384390765357004]
  We propose an integrated federated split learning and hyperdimensional computing framework for emerging foundation models. This novel approach reduces communication costs, computation load, and privacy risks, making it suitable for resource-constrained edge devices in the Metaverse.
  arXiv  Detail & Related papers  (2024-08-26T17:03:14Z)
• Spiker+: a framework for the generation of efficient Spiking Neural Networks FPGA accelerators for inference at the edge [49.42371633618761]
  Spiker+ is a framework for generating efficient, low-power, and low-area customized Spiking Neural Networks (SNN) accelerators on FPGA for inference at the edge. Spiker+ is tested on two benchmark datasets, the MNIST and the Spiking Heidelberg Digits (SHD)
  arXiv  Detail & Related papers  (2024-01-02T10:42:42Z)
• Flash-LLM: Enabling Cost-Effective and Highly-Efficient Large Generative Model Inference with Unstructured Sparsity [12.663030430488922]
  We propose Flash-LLM for enabling low-cost and highly-efficient large generative model inference on high-performance Cores. At SpMM kernel level, Flash-LLM significantly outperforms the state-of-the-art library, i.e., Sputnik and SparTA by an average of 2.9x and 1.5x, respectively.
  arXiv  Detail & Related papers  (2023-09-19T03:20:02Z)
• Latency-aware Unified Dynamic Networks for Efficient Image Recognition [72.8951331472913]
  LAUDNet is a framework to bridge the theoretical and practical efficiency gap in dynamic networks. It integrates three primary dynamic paradigms-spatially adaptive computation, dynamic layer skipping, and dynamic channel skipping. It can notably reduce the latency of models like ResNet by over 50% on platforms such as V100,3090, and TX2 GPUs.
  arXiv  Detail & Related papers  (2023-08-30T10:57:41Z)
• DeepGEMM: Accelerated Ultra Low-Precision Inference on CPU Architectures using Lookup Tables [49.965024476651706]
  DeepGEMM is a lookup table based approach for the execution of ultra low-precision convolutional neural networks on SIMD hardware. Our implementation outperforms corresponding 8-bit integer kernels by up to 1.74x on x86 platforms.
  arXiv  Detail & Related papers  (2023-04-18T15:13:10Z)
• An Adaptive Device-Edge Co-Inference Framework Based on Soft Actor-Critic [72.35307086274912]
  High-dimension parameter model and large-scale mathematical calculation restrict execution efficiency, especially for Internet of Things (IoT) devices. We propose a new Deep Reinforcement Learning (DRL)-Soft Actor Critic for discrete (SAC-d), which generates the emphexit point, emphexit point, and emphcompressing bits by soft policy iterations. Based on the latency and accuracy aware reward design, such an computation can well adapt to the complex environment like dynamic wireless channel and arbitrary processing, and is capable of supporting the 5G URL
  arXiv  Detail & Related papers  (2022-01-09T09:31:50Z)
• Reconfigurable co-processor architecture with limited numerical precision to accelerate deep convolutional neural networks [0.38848561367220275]
  Convolutional Neural Networks (CNNs) are widely used in deep learning applications, e.g. visual systems, robotics etc. Here, we present a model-independent reconfigurable co-processing architecture to accelerate CNNs. In contrast to existing solutions, we introduce limited precision 32 bit Q-format fixed point quantization for arithmetic representations and operations.
  arXiv  Detail & Related papers  (2021-08-21T09:50:54Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.