During the 2023 Service Trade Fair, Zhongke Yushu launched the latest self-developed high-performance network "sharp tool" - KPU FLEXFLOW®-2100R RDMA acceleration DPU card. The release of this product marks Zhongke Yushu's continuous innovation in the fields of high-performance computing and data centers. It aims to build a "super highway" for internal communication of computing clusters for high-speed networks and high-performance storage, and assist the field of high-performance computing. Innovation.
Standing in the wave of the digital age, the demand for data transmission efficiency in high-performance computing and data center networks is a potential computing power bottleneck that is easily overlooked at present. RDMA technology that provides higher bandwidth and lower latency network performance has emerged as the times require. born.
KPU FLEXFLOW®-2100R RDMA acceleration DPU card is the first DPU accelerator card with RDMA function independently developed by Zhongke Yushu focusing on the new computing power infrastructure of DPU. It integrates high performance, stability, convenience and versatility, provides 2x100GbE network port connection capabilities, supports RoCEv2 hardware offloading capabilities and lossless network capabilities, and can be widely used in high-performance computing, storage, big data, and AI /ML and other fields, providing users with flexible and high-performance solutions.
Extremely fast—industry-leading throughput and latency performance
KPU FLEXFLOW®-2100R RDMA acceleration DPU card is based on the highly scalable RDMA engine with independent intellectual property rights of Zhongke Yushu. It supports high throughput and low-latency direct memory access, achieving true extreme speed in high-performance computing and data center networks. transmission.
High throughput: According to actual testing, KPU FLEXFLOW®-2100R has excellent throughput performance in send, read, and write test scenarios; especially in large data packet tests, the throughput performance can basically reach line speed. It can be seen that the performance of KPU FLEXFLOW®-2100R reaches or exceeds the mainstream level in the industry in scenarios of data packets of different specifications.
Low latency: According to actual measurements, KPU FLEXFLOW®-2100R's latency data in test scenarios of sending, reading, and writing small files below 4K are all below 5us, with a minimum of 3us, which is better than the performance level of mainstream RDMA smart network cards at home and abroad.
Stability - ensuring the reliability of operation, maintenance and data transmission
Stability is also one of the focuses of KPU FLEXFLOW®-2100R product development. In order to improve the stability of the product, Zhongke Yushu focuses on four key technical aspects: selective retransmission, out-of-order concurrent execution, real-time system controllability and QP-level precise exception management. This multi-dimensional ensures the high-performance computing and data In the central network, RDMA products can provide excellent stability, ensure that data transmission is controllable and operational, and that data packets are not lost.
Convenience - Easily implement high-performance network development and deployment
KPU FLEXFLOW®-2100R provides users with a convenient and efficient development and deployment experience, helping them achieve high-performance network deployment more easily. HADOS®, a software development platform completely independently developed by Zhongke Yushu, integrates a rich set of RDMA Verbs APl. By calling the API, the upper-layer application software can quickly utilize RDMA resources, accelerate the development and integration of RDMA applications, and greatly save developers' time. and energy.
Universal - ideal choice for multiple industries and scenarios
KPU FLEXFLOW®-2100R also offers wide versatility. It not only supports Multi-Host, realizes virtualization, pooling and sharing, and is suitable for a variety of deployment environments from data centers to cloud computing, but also supports the most widely used congestion control algorithm DCQCN to ensure that network traffic is controlled under congestion conditions. Stay highly stable and controllable. Most importantly, the product provides complete network functions, making it adaptable to multiple application scenarios from high-performance computing, storage to virtualization.
Under the national east-west computing strategy and the wave of ever-changing new technologies, the computing power base is evolving towards ultra-high bandwidth and ultra-low latency. In the future, Zhongke Yushu will also focus on new infrastructure, continue to explore the depth and breadth of computing, network, storage, security and other technologies, increase efforts to promote the productization and application of DPU software and hardware technology, and work with partners to lead a new generation The development of data center computing architecture lays a solid foundation for the era of digital economy!