Extreme Designs of Inspur Servers to Drive Innovation

28 April 2020

As computing transitions from traditional computing technologies to emerging intelligent computing technologies, such as cloud computing, big data, and artificial intelligence, etc., it is evident that data centers have become larger and more specialized, and data-rich applications are also being released at a faster pace. As a basic computing device, a server needs to be able to adapt to changes and jump out of the box of traditional design ideas.

While focusing on the five RASUM (reliability, availability, serviceability, usability and manageability) features of traditional servers, Inspur's servers have been fundamentally upgraded in design, form and function to deliver the ultimate hardware design for application scenarios in terms of optimization and form factor, and provide a suitable platform for cloud computing, AI, edge applications and other application scenarios.

Extreme Design for Application Scenarios

In response to cutting-edge application scenarios, such as AI, cloud computing, and edge applications, Inspur products push physical boundaries in terms of density, performance, storage density, and I/O capacity, and have achieved a seamless fit between applications and their deployment environments.

Inspur's AGX-2 and AGX-5 offer users with industry-leading density and performance for AI.

  • AGX-2 (NF5288M5): A Unique 2U, 8 V100, NVLink 2.0 AI Server with the Highest Computing Density

AGX-2 supports the deployment of 8 NVIDIA® Tesla® P100 GPUs with NVLink or PCI-E interfaces in a 2U space, and is the first in the industry to support NVLink 2.0 and the newly released NVIDIA® Tesla® V100 with an interconnect bandwidth between GPUs of up to 150GB/s. The server has good I/O expansion capability, providing 8 NVMe/SAS/SATA hot-swappable hard drives, which can support up to 4 high-speed cluster interconnects of 100 Gbps EDR InfiniBand interface cards. The AGX-2 not only supports conventional air cooling, but also on-chip water cooling to provide a higher computing energy efficiency ratio.

  • AGX-5 (NF5888M5): the most powerful AI super server

Inspur's AGX-5 server is configured with 16 of the currently most powerful NVIDIA Tesla® V100 Tensor Core 32GB GPUs, which support the next generation of more powerful GPU accelerators, providing 2 PFlops of single-server AI computing performance. Compared with previous GPU servers of the same type, the AGX-5 can support 4 times the  hyperscale deep neural network model and increase the training speed tenfold.

The AGX-5 uses the industry's most advanced NVIDIA NVSwitch™ interconnect structure, which is based on NVIDIA's latest HGX-2 platform and can achieve a 48-channel, 2.4TB/s full chip group high-speed interconnect and 512GB HBM2 global shared ultra-high-speed graphics cache, to provide a nearly linear AI calculation performance ratio. In addition, Inspur’s AGX-5 is equipped with 2 28-core powerful CPUs to provide top-level general-purpose computing performance, 6 TB of persistent memory that can provide high-speed access to large data, and symmetrical balanced design that supports the upgrade of dual physical partitions.

For cloud computing scenarios, Inspur's signature open computing products, Crane Mountain, NF8380M5, and i48, are designed for hyper scale high-density cloud data centers.

  • Crane Mountain (NF8260M5) - OCP Accepted first ever high-density 4-socket server optimized for cloud computing scenarios

The NF8260M5 can support 4 Intel Xeon Scalable processors, 48​DIMM slots, and 8 PCI-e 3.0, 1 OCP slot, and can be expanded to up to 4 GPU accelerators. In terms of storage, the flexible combination of 6 U.2 NVMe SSDs and 24 2.5-inch mechanical hard drives forms an efficient tiered storage system.

The NF8260M5 has been optimized in energy consumption, density, operation and maintenance management, and TCO, etc. Compared with traditional servers, the NF8260M5 has doubled compute density and lowered power consumption by 10%. At the same time, when used for a hyperscale data center, this can double the operation and maintenance management efficiency and make the management network and data network more concise.

This product can also bring considerable benefits to cloud computing users in terms of TCO, which makes it possible for users to obtain an ideal balance between performance and cost. According to statistics, their hardware costs can be reduced by 7% - 12%, and operating costs can be reduced by 5% - 7%.

  • NF8380M5 - the OCP community's first 4-socket server based on Olympus specifications that is designed specifically for CSP users and hyperscale data centers

The NF8380M5 server can be expanded to include 4 Skylake and Cascade Lake processors, and 48 memory slots. Inspur fully considered the requirements of special application scenarios when designing this product and adopted a super-powered VR power supply module to increase the maximum power consumption of a single processor from 205W to 255W, and enable 24 memory slots to support Optane memory.

This server has excellent large-scale transaction and data analysis processing capabilities. Through an external GPU expansion box, it can meet all the offline training needs of AI models from the small scale to the hyperscale.

Through configuration changes, this versatile 4-socket server can provide a highly suitable platform for a variety of business applications. It is not only suitable for the businesses of Internet companies, but also satisfies the application needs of traditional enterprises and government users.

  • i48 - a high-density server that conforms to the Scorpio multi-node server standard set by the ODCC and meets the needs of new high-density data centers

The i48 uses an innovative design that combines the advantages of blade servers and rack servers. It is able to support 8 high-density compute nodes in a 4U standard chassis, which increases the deployment efficiency by 8 times compared with the traditional rack server cluster deployment method. It also reduces power consumption by more than 18% through unified power supply and cooling.

The micro-cabinet design and component optimization adopted by the i48 can reduce the cost of a single node by 8% compared to a blade server and by 15% compared to a typical rack server.

It exhibits the features of high-density, high-efficiency, reliability, and intelligence perfectly in a limited space. It is suitable for cluster applications such as cloud computing and high-performance computing clusters, and is especially suitable for users in the telecom and Internet industry, and large and medium-sized enterprises that have strict requirements for servers.

For edge computing, Inspur's compact and powerful NE5260M5 is specially designed for AI at the edge and capable of meeting all edge application scenario requirements.

  • NE5260M5 - compliant with the ODCC OTII standard and is specially designed for various edge computing applications in the 5G era

The NE5260M5 is suitable for edge data center deployment, and can undertake various network edge applications such as NFVI, video, AI, and CDN, etc. It has a height of 2U, width of 19 inches, and depth of 430mm, which is only about half the depth of a traditional standard server. This product adopts Intel Scalable processors, which can be configured with 2 processors and 16 memory slots, 2 of which support Optane Memory. The motherboard integrates 2 10G SFP network cards, and 6 PCIe-3.0 interfaces. In terms of storage, it can support 6 HDD/SSDs and 2 2.5-inch M2 SSDs. It can support 2 NVIDIA® Tesla® V100 accelerator cards or 6 NVIDIA® Tesla® T4 accelerator cards to support edge AI workloads.

The deployment environment of edge computing is relatively harsh, and the location may be in villages, roadsides, electrical towers, etc. The NE5260M5 can withstand high temperatures and high humidity for a long time, and is dust-proof and shock-proof. It has robust environmental adaptability and is specially designed for the special environment of an edge computing data center. In response to the absence of dedicated operation and maintenance personnel, Inspur has optimized the design of the NE5250M5 to include a self-repair mechanism through remote online maintenance and management.

The Ultimate Hardware Design

The ultimate objective in the logical design of the server is to achieve maximum application efficiency, and the ultimate objective in the physical design is to achieve a precise balance between performance, space, and energy efficiency. Inspur’s servers are supported by a series of technologies, such as structural strength, efficient heat dissipation and power supply, and the ability to deliver maximum performance, energy efficiency and scalability in a specific space.

High-strength structural design: scalability is continuously improved under the premise that the height is constant. For example, if the NF5486M5 storage server supports 100 3.5” hard disks in a 4U space, the massive weight will cause deformation to the server motherboard and cabinet, and the hard disk rotation will produce severe resonance. In order to solve these problems, Inspur’s servers adopt leading technologies, such as three-dimensional layering, sub-unit reinforced frame, and resonance shielding module, to provide structural strength capable of ensuring the stability of the system.

Extreme heat dissipation design: the increase in space density directly increases power consumption in the unit space. Higher component density complicates airflow, especially in high-power components such as GPUs, which can easily cause overheating and downtime. For example, in the case of the NF5486M5 storage server, the hard drives of this product are divided into 8 rows. When the server is running, each layer of hard drives will cause preheating of about 2°C to the hard drives behind them. If the server is deployed in a data center with an ambient temperature of 35°C, the operating temperature of the hard disks at the very end of the NF5486M5 will be very close to the critical temperature. In the design process, Inspur introduced a precise air distribution technology and ventilation in the partition layering to reduce the hard disk preheating to less than 1 degree, so as to ensure that the server can also operate normally in high temperature environments.

Extreme power supply design: a good power supply design is a prerequisite for the stable operation of a server. As space density increases, so does the power supply in the unit space, and resulting problems such as voltage fluctuations, electromagnetic interference and signal distortion will worsen. For example, the power of the NF5288M5 at full configuration exceeds 3KW, which is more than 6 times that of a general-purpose 2U server. Inspur uses ultra-low-loss power transmission technology and high-efficiency VR designs, to achieve over 97% power supply efficiency, reduce power transmission loss and solve problems like electromagnetic interference.

According to Gartner, x86 server shipments and vendor revenues in 2019 were 12.497 million units and USD 69.36 billion, down 3.2% and 1.8% respectively year-on-year. In contrast, Inspur has maintained its growth momentum. Inspur is ranked among top three companies in the world both in server market share and revenue, with a 7.7% increase in sales, making Inspur the manufacturer with the highest growth rate among the world's top five manufacturers. With 10.3% market share in 2019, Inspur’s server designs are demonstrably fulfilling market needs for extreme hardware. . Inspur will continue to integrate leading technology into its products to provide users with products that deliver the most cutting-edge features.