_version_ 1866918065378164736
author Zuo, Pengfei
Lin, Huimin
Deng, Junbo
Zou, Nan
Yang, Xingkun
Diao, Yingyu
Gao, Weifeng
Xu, Ke
Chen, Zhangyu
Lu, Shirui
Qiu, Zhao
Li, Peiyang
Chang, Xianyu
Yu, Zhengzhong
Miao, Fangzheng
Zheng, Jia
Li, Ying
Feng, Yuan
Wang, Bei
Zong, Zaijian
Zhou, Mosong
Zhou, Wenli
Chen, Houjiang
Liao, Xingyu
Li, Yipeng
Zhang, Wenxiao
Zhu, Ping
Wang, Yinggang
Xiao, Chuanjie
Liang, Depeng
Cao, Dong
Liu, Juncheng
Yang, Yongqiang
Bai, Xiaolong
Li, Yi
Xie, Huaguo
Wu, Huatao
Yu, Zhibin
Chen, Lv
Liu, Hu
Ding, Yujun
Zhu, Haipei
Xia, Jing
Xiong, Yi
Yu, Zhou
Liao, Heng
author_facet Zuo, Pengfei
Lin, Huimin
Deng, Junbo
Zou, Nan
Yang, Xingkun
Diao, Yingyu
Gao, Weifeng
Xu, Ke
Chen, Zhangyu
Lu, Shirui
Qiu, Zhao
Li, Peiyang
Chang, Xianyu
Yu, Zhengzhong
Miao, Fangzheng
Zheng, Jia
Li, Ying
Feng, Yuan
Wang, Bei
Zong, Zaijian
Zhou, Mosong
Zhou, Wenli
Chen, Houjiang
Liao, Xingyu
Li, Yipeng
Zhang, Wenxiao
Zhu, Ping
Wang, Yinggang
Xiao, Chuanjie
Liang, Depeng
Cao, Dong
Liu, Juncheng
Yang, Yongqiang
Bai, Xiaolong
Li, Yi
Xie, Huaguo
Wu, Huatao
Yu, Zhibin
Chen, Lv
Liu, Hu
Ding, Yujun
Zhu, Haipei
Xia, Jing
Xiong, Yi
Yu, Zhou
Liao, Heng
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2506_12708
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Serving Large Language Models on Huawei CloudMatrix384
Zuo, Pengfei
Lin, Huimin
Deng, Junbo
Zou, Nan
Yang, Xingkun
Diao, Yingyu
Gao, Weifeng
Xu, Ke
Chen, Zhangyu
Lu, Shirui
Qiu, Zhao
Li, Peiyang
Chang, Xianyu
Yu, Zhengzhong
Miao, Fangzheng
Zheng, Jia
Li, Ying
Feng, Yuan
Wang, Bei
Zong, Zaijian
Zhou, Mosong
Zhou, Wenli
Chen, Houjiang
Liao, Xingyu
Li, Yipeng
Zhang, Wenxiao
Zhu, Ping
Wang, Yinggang
Xiao, Chuanjie
Liang, Depeng
Cao, Dong
Liu, Juncheng
Yang, Yongqiang
Bai, Xiaolong
Li, Yi
Xie, Huaguo
Wu, Huatao
Yu, Zhibin
Chen, Lv
Liu, Hu
Ding, Yujun
Zhu, Haipei
Xia, Jing
Xiong, Yi
Yu, Zhou
Liao, Heng
Distributed, Parallel, and Cluster Computing
Artificial Intelligence
Hardware Architecture
Machine Learning
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.
title Serving Large Language Models on Huawei CloudMatrix384
topic Distributed, Parallel, and Cluster Computing
Artificial Intelligence
Hardware Architecture
Machine Learning
url https://arxiv.org/abs/2506.12708