Introduction: Storage Innovation in the PCIe 5.0 Era

With the rapid development of large model training, edge computing and hyper-scale data centers, storage systems are facing multiple challenges in terms of performance, efficiency and reliability. The recently released outstanding performance data of Union Memory PCIe 5.0 enterprise solid-state drives (ESSDs) UH812a/UH832a has attracted much attention. Its sequential write speed is 69% higher than that of international competitors, its random write speed is 52% higher than that of international competitors, and its sequential read and random read data rank Top1 among international competitors. H3C UniServer G7 is a new series of computing servers developed by H3C Group, widely used in computing scenarios such as virtualization, AI computing, scientific computing and data warehouse analysis. How will the alliance of two powerful enterprises between Union Memory and H3C bring about a leap in the performance of the storage systems?

This article will, based on the latest test results of SPECstorage Solution 2020, deeply analyze the outstanding performance of Union Memory PCIe 5.0 ESSD when used with H3C UniServer G7 servers, and showcase its leading position in enterprise core business, mixed loads of data centers, and AI scenarios.

Test Background and Environment

Test Standard

SPECstorage Solution 2020, launched by the Standard Performance Evaluation Corporation, is an authoritative industry-wide performance test standard for the storage systems. It simulates real-life enterprise-level storage workloads and is used to comprehensively and objectively evaluate the performance of storage systems in different scenarios. Through this test, enterprise users can accurately measure the maximum sustainable throughput offered by different storage solutions, enabling them to precisely match their storage systems with actual business requirements and avoiding performance bottlenecks.

Tested Products

This test used H3C UniServer R3900 G7 and H3C UniServer R3950 G7 servers equipped with Union Memory UH832a.

H3C UniServer R3900 G7 is a 2U single-socket server based on Intel's new generation Birch Stream platform, designed for general computing scenarios. H3C UniServer R3950 G7 is an enterprise-grade 2U single-socket server powered by AMD EPYC 9005 series processors, delivering both high performance and cost efficiency. Union Memory UH832a, with its advanced technology and excellent quality, provides efficient data storage support for the servers.

Key Test Results

1. AI Image Processing Test - Ranking Top1

This test simulates the AI TensorFlow image processing workload, involving high concurrent reading/writing on a large amount of unstructured data (e.g. images and videos). Typical application scenarios include AI training/ inference and computer vision tasks, which require high throughput and high IOPS.

In this test, two H3C servers equipped with UH832a demonstrated outstanding AI workload processing capabilities, achieving overall response times (ORT) of 0.35 milliseconds and 0.24 milliseconds, respectively, surpassing the data of the H3C UniServer R4900G6 Ultra server equipped with the previous generation PCIe 4.0 ESSD and domestic competitor Gen5 SSD (both 0.47 milliseconds), and the indicators were improved by 34.29% and 48.94%, respectively.

Fig. 1 Comparison of AI Image Processing Test Results (1)

Among them, the H3C UniServer R3950 G7 configured with 6.4TB UH832a delivers AI workload performance on par with the Super Micro Computer AS-1115CS-TNR server on the list, achieving top ranking in overall response time. This server uses the same AMD EPYC 5th 9355P processor as the H3C UniServer R3950 G7, and is equipped with a Samsung PM1743 3.84TB SSD. During the test, both products were able to handle 75 AI tasks, achieving identical throughput of 7337 MB/s, and matching ORT of 0.24 milliseconds.

Fig. 2 Comparison of AI Image Processing Test Results (2)

The test results show that when Union Memory PCIe 5.0 ESSD is combined with the H3C UniServer G7 server, it can provide high-performance computing and storage capabilities for AI workloads, achieving ultra-low latency and top-tier benchmark rankings.

2. Electronic Design Automation Testing - Benchmark Performance

This test simulates the typical behavior of an EDA (Electronic Design Automation) application. The program includes software tools and workflows used for designing semiconductor chips. Its workload is characterized by high concurrency and computing-intensive processes, which severely tests storage performance. This test involves large volume of metadata operations of small files and partial sequential reading/ writing of large files. Typical applications include semiconductor design, electronic engineering, and other scenarios that require low-latency metadata processing and high-bandwidth storage.

In this test, the two H3C servers equipped with UH832a achieved ORT of 0.04 milliseconds and 0.03 milliseconds for handling EDA mixed workloads, both surpassing the data of the H3C UniServer R4900G6 Ultra server with the previous generation PCIe 4.0 ESSD and domestic competitors' Gen5 SSD (0.07 milliseconds and 0.05 milliseconds). That is, compared to the data with the previous generation of storage device, the ORT indicator of Union Memory UH832a has increased by up to 57.14%, and this result also equalizes the performance of the Super Micro Computer AS-1115CS-TNR server.

Fig.3 Comparison of Electronic Design Automation Test Results

The test results show that Union Memory PCIe 5.0 ESSD UH832a has an extremely outstanding performance when handling mixed workloads characterized by massive small-file operations. It has a strong concurrent processing capability and can handle various types of I/O operations simultaneously, meeting the complex requirements of enterprise-grade applications and setting a new industry benchmark.

3. Genomic Pipeline Benchmark Testing - No.1 in Performance Rankings

This test models the entire processing pipeline of the genomic workflows, emphasizing the capability to quickly analyze and retrieve large-scale datasets. This type of workload is prevalent in bioinformatics, precision medicine and other scenarios that require efficient processing of massive genomic datasets.

In this test, the H3C UniServer R3900 G7 server equipped with UH832a completed the processing of 250 genomic analysis operations, with a throughput of up to 21236 MB/s and a total response time of 0.07 milliseconds. This result surpassed the previous data of the H3C UniServer R4900G6 Ultra server (0.10 milliseconds), further breaking the record on this list.

Fig. 4 Comparison of Genomic Pipeline Benchmark Testing Results

The test results demonstrate that Union Memory PCIe 5.0 ESSD can exhibit extremely high performance and ultra-low latency in scenarios involving massive data analysis and processing, providing strong support for real-time data processing.

4. Software Build Testing - Another Breakthrough

This test simulates the build operation in the software development environment, covering various scenarios such as Git code warehouse, CI/CD pipeline (Jenkins), and compilation build (GCC, LLVM). This type of workload is characterized by intensive metadata and typically involves a large number of frequent file reading/writing operations.

In this test, the combination of UH832a and H3C UniServer R3900 G7 servers achieved a stable throughput of 2440 MB/s and an ORT of 0.02 milliseconds. The R4900G6 Ultra equipped with the previous generation product demonstrates an ORT of 0.03 milliseconds, proving that both PCIe 4.0 and PCIe 5.0 ESSDs, when handling IOPS- intensive small-file random workloads, deliver outstanding throughput and low latency performance. PCIe 5.0 ESSD is even better in low latency performance. Especially in high concurrent access scenarios, the stability and high efficiency of Union Memory ESSD have been fully verified.

Fig. 5 Comparison of Software Construction Test Results

5. Video Data Acquisition Test - Debuts as Leader

This test simulates the storage load of Video Data Acquisition (VDA) and processing, such as security cameras, scientific experiment data acquisition, broadcasting, film and television production shooting, or real-time streaming media processing. These scenarios require the storage system to be capable of delivering high throughput, low latency sequential writing, as well as long-term stable performance.

This marks the inaugural benchmark submission by Union Memory in collaboration with H3C servers. Two H3C servers equipped with UH832a each processed 1,000 video streams, with approximate throughput data (4605 MB/s and 4608 MB/s), and the overall response times were 1.02 milliseconds and 0.92 milliseconds, respectively. Among them, the performance of the H3C UniServer R3950 G7 server equipped with UH832a once again equalized the Super Micro Computer AS-1115CS-TNR server, tied at the Top1.

Fig. 6 Comparison of Video Data Acquisition Test Results

The test results demonstrate that Union Memory UH832a provides strong support for video processing, real-time data acquisition and other operations. The processing of one thousand data streams indicates that the storage system has a high throughput capability when dealing with multiple video streams, while the reduced ORT ensures the real-time data writing, eliminating video frame loss or latency.

Performance Highlights: PCIe 5.0 Technology Unleashes Extraordinary Performance

Union Memory UH812a/UH832a uses Union Memory's high-performance enterprise-grade controller and the latest PCIe 5.0 high-speed interface, featuring extremely high bandwidth and ultra-low latency. It delivers over 2 times higher performance when compared to the previous generation PCIe 4.0 products, with 43% lower QD1 latency. Its key advantages include:

•       Ultra-high-speed Bandwidth: Double the bandwidth in comparison with PCIe 4.0, support single and dual ports and NVMe 2.0 protocol. The powerful I/O capability can meet the requirements of high-throughput scenarios such as AI training and inference, as well as real-time analysis, and accelerate data processing efficiency;

•       Ultra-low Latency: Targeting complex simulation programs and processing of massive datasets, the product delivers microsecond level average read/write latency, with <55μs latency in 4K random read at QD1, providing robust support to latency-sensitive data-intensive operations such as OLTP databases and virtualization;

•       Full-scenario High Compatibility: Diversified combinations of capacities and forms, support various capacities ranging from 1.6TB to 15.36TB, compatible with U.2 and E3.S forms, flexibly adapt to the storage requirements of mixed service scenarios from central cloud/edge cloud to hyper-scale data centers.

With the performance advantages of PCIe 5.0, UH812a/UH832a can be widely applied in various mainstream intelligent computing scenarios, such as:

•       AI and Big Data: Building large-scale storage systems to support hundred-billion-parameter model training, efficiently handling PB-level datasets;

•       Cloud and Edge Computing: Providing low-latency storage resource pools for hybrid cloud environments, accelerating the real-time decision-making process of edge nodes;

•       Financial Core Systems: Ensuring sub-millisecond response in high-frequency transaction and risk analysis scenarios, while PCIe 5.0 ecosystem compatibility further reduces the cost of architecture upgrades.

Conclusion: Driving the Evolution of Intelligent Computing through Storage Innovation

Union Memory PCIe 5.0 ESSD, based on technological breakthroughs, has demonstrated excellent performance on H3C general-purpose computing servers, fully verifying its leading position in the field of enterprise-grade storage. Through the authoritative tests of SPEC, Union Memory ESSD has redefined the performance benchmark of enterprise-grade storage, demonstrating its outstanding capabilities in high performance computing and broad ecosystem compatibility. In the future, Union Memory will continue to focus on the innovation of flash memory technologies, collaborate with ecosystem partners, and provide customers with higher-quality storage solutions, helping enterprises easily address the increasing challenges of data storage and fully unleash the value of intelligent computing!