Meet the complex computing needs of the AI ​​era. Review of Adam Server G952N6

The explosive development of generative AI and LLM has promoted the in-depth application of intelligent technology in thousands of industries. Accelerated computing has become the growth driver of IT infrastructure, especially the continuous iteration of products or technologies such as AI servers, CXL, and HBM. The excellent performance provided allows people to see its broader market opportunities. The content that needs to be rewritten is: EVOC Adam server G952N6 is a 4U10 GPU card dual-channel heterogeneous computing server designed by Ewanke equipped with Intel's third-generation Xeon scalable series processors. It is suitable for the Internet and IDC (Internet Data Center) ) Businesses such as cloud computing, big data, universities, artificial intelligence, and high-performance computing can meet the needs of scenarios such as large data scale, high application computing power, and flexibility.

From the appearance, the Adam Server G952N6 adopts a simple and elegant design style, which is easy to install, disassemble and maintain. This is a 4U 19-inch rack-mounted server with a net weight of 60.35KG and a size of 840mm ( Length) * 460mm (width) * 175mm (height), the recognizable logo on the front panel and the orange lines highlight the temperament of a big manufacturer.

The content that needs to be rewritten is: EVOC Adam server G952N6

The upper right corner of the front panel provides 2 USB 3.2 and 1 VGA interface, the upper left corner is a UID button/indicator, an RST button, a power switch button/indicator, and a health The status light, 1 hard disk indicator light and 2 LAN indicator lights allow users to see it clearly at a glance.

At the same time, G952N6 has 12 3.5-inch hot-swappable hard drive bays and can support 4 U.2 NVMe hard drives, which can fully meet the storage needs of application data such as generative AI, HPC, and big data analysis.

The content that needs to be rewritten is: EVOC Adam server G952N6

The G952N6 motherboard is equipped with 2 M.2 interfaces for system disks and provides up to 13 PCIe expansion slots to meet various expansion needs of customers. It supports up to 10 full-height, full-length, double-width 300W GPU accelerator cards. , suitable for deep learning, big data analysis and other applications. The board provides an IPMI dedicated GbE and an OCP 3.0 interface for flexible selection.

In operation and maintenance, you can check the operating status of the equipment through the BMC Web management interface, fault diagnosis LED, etc., mark the faulty machine through the UID indicator light of the mounting ear, quickly locate the faulty equipment, solve the problem in a timely manner, and simplify the maintenance work. , improve system availability. In addition, the G952N6 server also integrates an intelligent heat dissipation management system, which can intelligently adjust the heat dissipation mode according to the operating status of the entire machine, improve the energy efficiency ratio of the equipment, and save electricity costs for customers

The specific parameter configuration of the Yiwanke G952N6 server is as follows: - Processor: 2.5GHz eight-core Intel Xeon processor - Memory: 32GB DDR4 RAM -Storage: 1TB SSD - Network: Gigabit Ethernet interface - Operating system: Windows Server 2019 - Expansion slots: 4 PCIe x16 slots - Power supply: Dual redundant hot-swappable power supplies - Size: 2U rack-mounted design The above is the specific parameter configuration of the Yiwanke G952N6 server

Yiwanke G952N6 server parameter configuration

According to the above related configurations, a single CPU of this server has up to 40 cores and 80 threads, a maximum TDP of 270W, a maximum main frequency of 3.6GHz, and three sets of 11.2 GT/s UPI interconnection links. Paired with Intel C621A server chipset and 32 DDR4 3200 MTS memory. At the same time, it also supports 16 BPA PMEM type memories, effectively increasing the memory capacity

In the testing process, we conducted a preliminary inspection of the appearance structure of the G952N6 server, and separately inspected details such as on-board screen printing, OCP slot screen printing, serial port screen printing, hard disk light screen printing, PCIE slot, power light/switch screen printing, etc. , exquisite workmanship, clear silk screen, worthy of the quality of a major manufacturer.

Then we analyzed the server panel, OS installation, CPU function, memory function, PCH SATA disk, RAID/HBA card SATA/SAS disk/NVME disk, PCIE U.2 NVMe disk or onboard M.2 NVMe disk, GPU Functions, PCIE slots, onboard network ports, VGA interfaces, serial ports, USB interfaces, power interfaces, fan interfaces, OCP slots, etc. were tested for basic functions, and all test items were passed with excellent performance.

For example, when testing the PCIE slot, run the stress test of the PCIE card for 10 minutes. There are no error messages in dmesg and bmc sel, and there are no abnormalities in the PCIE card. The PCIE card can be recognized normally when the machine is turned on and off 15 times. The performance is excellent.

Test configuration table

Entering the performance test session, we conducted comprehensive tests on CPU performance, memory performance, hard disk performance, and graphics card performance.

First we perform a performance test on the CPU. With the help of the Linpack test program, we test the CPU's computing power by calculating the solution to a system of double-precision linear equations. The test results show that the measured values ​​are better than the theoretical values. The better floating-point computing performance allows the server to have excellent performance in engineering applications or programming environments, such as experimental simulations of physics or numerical calculations in computer programming. .

The content that needs to be rewritten is: Central Processing Unit (CPU) performance test

In the G952N6 server memory test, we chose two tools, STREAM (v5.10 version) and MLC (v3.9 version), for testing. Among them, STREAM is a mainstream memory bandwidth test program that has low requirements on CPU computing power, but puts greater pressure on CPU memory bandwidth. Through testing, we can get the sustainable operating maximum value of memory bandwidth, not the theoretical value. This program mainly uses four array operations, including array copy (Copy), array scale transformation (Scale), array vector sum (Add), and array composite vector sum (Triad). Array values ​​used in tests are in double precision

To test memory performance

Generally speaking, when applications obtain data from the processor cache or memory subsystem, they will encounter various latency issues. Coupled with memory bandwidth limitations, these delays can have an impact on application performance. MLC can be used to test these delays. For example, you can query the results of the memory access latency instruction, which represents the free memory access latency matrix between or within nodes in ns.

To test memory performance

To test memory performance

Judging from the memory test results, the performance is generally higher than the theoretical value of the official website. It can not only handle common tasks such as data query, application, and call, but also play a greater role in popular high-performance computing and other applications. . As can be seen in the model configuration, ECC memory can implement error checking and correction, which to a large extent ensures the overall security and stability of the system and can minimize the defect rate, while RDIMM memory can reduce latency. , suitable for large-capacity, high-speed, real-time application scenarios. In addition, with the help of LRDIMM, the use of buffers on the data bus can reduce the electrical load of the system memory controller, allowing the server memory bus to reach higher The operating frequency greatly increases the capacity supported by the memory.

In the hard disk performance test, we tested SAS disks, NVMe disks and SATA disks, using FIO (v3.15 version) software. FIO (Flexible I/O Tester) is a professional software for testing Tools for disk, file system, block device and network device performance that can simulate different types of I/O loads, including random or sequential read and write, mixed read and write, random or sequential access, etc., and can be highly customized for testing configurations , so that the whole machine's support for hard disk performance can be fully demonstrated.

SAS disk sequential read and write test

The contents of the SAS disk RAID test need to be rewritten

SAS disk random read and write test

The content that needs to be rewritten is: NVMe disk sequential read and write test

Conduct random read and write tests on NVMe disks

SATA disk sequential read and write test

Random read and write test of SATA disk

SATA disk RAID test

The G952N6 server provides three hard drive options: SAS, SATA, and NVMe to meet various complex data storage needs and backup. After testing, we found that the actual performance of the hard drive is better than the theoretical value on the official website. For example, NVMe disks have the characteristics of high performance and low latency. They can make full use of the high bandwidth capabilities of PCIe to provide higher data transfer rates and IOPS, and are suitable for scenarios such as big data processing and high-performance computing. SAS hard drives perform well in RAID, providing higher data protection and fault response capabilities

We use CUDA tools to test GPUs. CUDA is a high-performance parallel computing platform that allows developers to use C/C, Fortran and other programming languages ​​​​to perform parallel computing on NVIDIA GPUs, providing powerful computing capabilities and efficient The data transmission method makes GPU widely used in scientific computing, machine learning, deep learning and other fields. CUDA-Z is a tool used to query graphics card information and test graphics card performance. It can conduct professional testing of NVIDIA GPUs. It is also the test software we selected this time (CUDA-Z 10.251).

The content of CUDA-Z FP32 test needs to be rewritten

CUDA-Z FP64 test is a tool for testing GPU double-precision floating point performance. It helps users evaluate how well their graphics card handles double-precision floating point operations. By running the CUDA-Z FP64 test, users can understand how well their graphics card performs when handling high-precision computing tasks. This is important for applications that require scientific computing, data analysis, and other applications that require high-precision calculations. By using the CUDA-Z FP64 test, users can choose the right graphics card to meet their specific needs and optimize their computing tasks

According to the test data, the actual test value of graphics card performance is better than the official theoretical value. Take FP32 and FP64 as examples. The former is the default accuracy adopted by most deep learning frameworks. It can provide faster calculation speed while ensuring a certain calculation accuracy. Single-precision floating point performance is high and can better meet the needs of 3D graphics processing, such as engineering modeling and design tasks. The high value of double-precision floating point number FP64 can meet the needs of precise scientific calculations, such as molecular modeling, fluid dynamics and other scenarios

In addition, we also tested the power consumption of the machine. The results show that the maximum power consumption of the G952N6 server is 3224 W when turned on, the no-load power consumption of the system is 823 W, and the system power consumption of 50% load ( Run PTU FIO gpu_burn under the system to let the CPU, memory, and NVMe disk/hard disk reach 50%, 10 minutes) is 3365 W. The maximum power consumption under the system (run YES DD under the system to let the CPU, memory, NVMe disk/hard disk reach 100% %, 10 minutes) is 3652 W. This server has good performance and its power consumption control and energy saving performance are also very commendable.

Summarize:

Through the above tests, the Adam Server G952N6 server performed very well in terms of processor and memory performance. Its multi-core and multi-thread capabilities can handle business peaks with ease when processing high concurrency and large-scale data. The expansion capabilities of PCIe can meet the current hot AI needs in any scenario, and can fully unleash the potential of the GPU to bring huge performance to applications such as deep learning and big data analysis. With the support of various hard disk forms such as NVMe, It can better meet various complex needs in business scenarios such as the Internet, AI, and cloud computing, such as large-scale complex data storage, elastic resource expansion, high-performance computing, etc. From the overall performance, the Adam Server G952N6 It is a very competitive product for the cloud and AI era.

The above is the detailed content of Meet the complex computing needs of the AI ​​era. Review of Adam Server G952N6. For more information, please follow other related articles on the PHP Chinese website!