AI server NF5488A5 ranks first in the MLperf’s single server performance

15 January 2021

1.What is an AI server?

The three elements of deep learning are data, algorithms, and computing power. Among them, data is the foundation, algorithms are tools, and computing power is the booster. The increase in computing power drives the development of deep learning. Deep learning has developed slowly before, except for algorithms. In addition to the reasons, a very important reason is the lack of computing power. The most important support for solving computing power is the AI server. From the perspective of the hardware architecture of the server, the AI server is a heterogeneous server. In the heterogeneous mode, different combinations can be used according to the scope of the application, such as CPU+GPU, CPU+TPU, CPU+other accelerator cards, etc. Compared with ordinary servers, there is no difference in memory, storage, and network. The main aspects are big data, cloud computing, artificial intelligence, etc., which require larger internal and external storage to meet the collection and sorting of various data. Ordinary GPU servers are generally single-card or dual-card. AI servers need to undertake a lot of calculations. Generally, more than four GPU cards are configured, and AI server clusters are even to be built.

2.About AI server NF5488A5

AI server NF5488A5 completed the ResNet50 model training in only 33.37 minutes, ranking first in single server performance of Resnet50 benchmark.NF5488A5 is a new generation of AI server developed by Inspur. It has the world's first single test performance of MLperf. and it supports 8 third-generation NVLink fully interconnected NVIDIA A100 GPUs in 4U space, equipped with 2 AMD EPYC 7742 processors that support PCIe4.0.AI server NF5488A5 provides 5 petaFLOPS of extreme stand-alone training performance and ultra-high data throughput, oriented to image and video , Voice recognition, financial analysis, intelligent customer service and other typical AI application scenarios.

3.Functions and features

First of all, the AI server NF5488A5 has the ultimate AI training performance. The entire machine uses PCIe4.0 communication and uses NVLink3.0 fully interconnected A100 GPU, which can provide up to 5 petaFLOPS AI computing power. Secondly, the AI server NF5488A5 has the ultimate communication rate. compared with the previous generation,PCIe4.0 communication between CPU and GPU can double the bandwidth, P2P communication between GPUs can reach 600GB/s, which can greatly reduce data delay. Thirdly, it’s hardware design is also the most superior, 4U space, modular design, suitable for a wider range of data center environments, which can greatly reduce deployment costs and improve operating efficiency. Finally, it can accelerate AI application innovation. The AI server NF5488A5 is suitable for various AI applications such as image and video, voice recognition, financial analysis, and intelligent customer service, providing powerful computing power support and accelerating application innovation.

AI server NF5488A5 was launched in May. It is a new type of AI server, driven by 8 NVIDIA A100 GPUs, fully interconnected with the third-generation NVLink, plus two latest AMD CPUs supporting PCIe 4.0. It provides users with ultimate AI performance and ultra-high-speed bandwidth, while providing powerful computing support for various AI scenarios, such as smart customer service, financial analysis, smart city and smart language processing.