Strength is recognized again! Tianyi Cloud won the 2024 DPU
2024-08-14
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
Recently, the 2024 DPU & AI Networking Innovation Conference with the theme of "Intelligently Driven Networks and Powering the Future" was held in Beijing. The conference recognized the units and projects that have made outstanding achievements in DPU and AI network technology innovation and practical applications. Tianyi Cloud Technology Co., Ltd. won the Innovation Engine Award, and "Zijin DPU Computing Power Offloading and Network Acceleration Application" won the Practice Pioneer Award. The technical innovation strength and application practice results have been recognized by the industry again. At the AI Computing Network Technology Forum, Fan Xiaoping, a senior R&D expert at Tianyi Cloud Technology Co., Ltd., delivered a speech and shared Tianyi Cloud's technical innovation in high-performance intelligent computing networks.
"Innovation Engine Award" Award Ceremony "Practice Pioneer Award" Award Ceremony
The demand for intelligent computing has increased dramatically in the era of artificial intelligence, and unprecedented requirements have been placed on the network. Fan Xiaoping said that building a high-performance intelligent computing network faces many challenges. At the terminal level, RDMA network cards need to access multiple network planes such as storage and intelligent computing parameter planes and face the problem of accelerating business integration. If RDMA network cards want to achieve extreme performance, they need to overcome the difficulty of high-performance communication libraries. At the network level, the training data and parameters of large AI models are huge, and the training involves tens of thousands of cards in parallel. This not only puts higher requirements on network performance, reliability, security, bandwidth, etc., but also requires the establishment of a large-scale RDMA network to support it.
Fan Xiaoping, senior R&D expert at Tianyi Cloud Technology Co., Ltd.
In order to meet the above challenges, Tianyi Cloud actively explores new technologies in the field of intelligent computing networks and builds a high-performance intelligent computing network that can be expanded to a cluster of 10,000 cards. The parameter-side RDMA network adopts a three-layer networking to achieve end-to-end collaboration, software-hardware integration, and business perception. In terms of RDMA network card optimization, Tianyi Cloud has developed the Zijin RDMA network card based on the Zijin DPU base, which has achieved four uses for one card and supports a programmable congestion control framework. In terms of congestion control, Tianyi Cloud has launched the CTCC congestion control algorithm, which can eliminate the complex waterline configuration of the switch and can select different preference strategies on different end sides, such as preferring high throughput or low latency. In terms of storage networks, Tianyi Cloud's three-stack fusion protocol stack SF-STACK supports dynamic selection of transport layer protocols, has the advantages of high performance and high reliability, shields hardware differences, and expands the types of deployable networks. In addition, Tianyi Cloud has launched a high-performance collective communication library CTCCL, which focuses on multi-path load balancing, fault detection and recovery, and can optimize network paths and ensure network availability.
At present, Tianyi Cloud's high-performance intelligent computing network supports VPC/object storage (VxLAN) access, provides parallel file storage (RoCE) access, and realizes high-performance storage engine LAVA docking through Zijin DPU, which can reduce network planes and reduce network complexity. With the advantages of a single card supporting access to multiple network forms and a single network carrying multiple transmission flows, Tianyi Cloud's high-performance intelligent computing network has achieved remarkable results in assisting intelligent computing and high-performance storage, and can help enterprises effectively reduce costs and improve efficiency.
As various industries continue to move to the cloud and use data, the synergy between networks and computing power will further promote the vigorous development of the digital economy. Tianyi Cloud will adhere to technological innovation, explore new intelligent computing network solutions, and inject strong momentum into the digital intelligence development of thousands of industries. (China.com)
