(observer Network News) May 14 news, Ali Cloud AI high-performance network architecture HPN 7PlayingbaccaratonlineThe. 0 result paper was included by SIGCOMM2024, which became the first paper on the network architecture of AI intelligent computing cluster in the history of SIGCOMM.
SIGCOMM is the most authoritative computer communication network meeting in the world. Google Jupiter network has developed into a classic architecture after it was selected into SIGCOMM. Some experts pointed out that Aliyun HPN7.0 is expected to become a new paradigm for the next generation of AI high-performance network architecture. It is reported that a total of six papers in Aliyun have been included by SIGCOMM this year.
Large model for AI infrastructurePlayingbaccaratonlineTo meet the higher requirements, Aliyun's new generation HPN7.0 architecture is developed for the super-large-scale and high-stability network requirements in the AI era. HPN 7.0innovatively designs the network architecture of "double up connection + multi-rail + double plane", and develops Solar-RDMA and ACCL communication library by ourselves, which can realize high performance and high stability interconnection of the network.
HPN7.0 Architecture: a High performance Network Cluster designed for AI
Starting from September 2023, HPN7.0 has launched a large-scale deployment in Aliyun. The training performance of the large model is 14.9% higher than that of the previous generation architecture in typical scenarios, and the overall stability of the intelligent computing network is greatly improved. Not long ago, Aliyun released Tongyi Qianwen 2.5 version of the big model, Chinese performance to catch up with and surpass GPT-4Turbo, is based on HPN7.0 high-performance network cluster training.
Cai Dezhong, head of infrastructure network at Aliyun, said: "Aliyun began to explore the predictable network of end-to-end network convergence in 2017, and HPN7.0 went a step further, extending the architecture of end-to-end network convergence from network protocol stack to network architecture and communication library, realizing a new network cluster architecture innovation for the AI intelligent computing era."
It is reported that SIGCOMM, which began in 1977, is the top international academic conference in the computer network field, with an annual admission rate of between 10% and 20%. The selected results are very likely to become the mainstream technology paradigm standard. Previously, Google's traditional data center network Jupiter paper SIGCOMM was included, which has become a classic architecture; this time Aliyun HPN7.0 architecture won SIGCOMM's first AI high-performance network paper, which is the first time that a Chinese company has provided a new high-performance standard for the design and practice of global AI network infrastructure.
Aliyun has been working in the field of high-performance network for many years, put forward the predictable network technology system of end-to-end network convergence, and took the lead in the large-scale practice of advanced technologies such as RDMA low-latency network and intelligent cluster network architecture HPN 7.0in the industry. Alibaba has also been rated as one of the top ten online research institutions in the AMiner and is the only Chinese science and technology company on the list.