AMD EPYC™ 9004 Series Processors Enable Transformative Flash Storage Performance

The PCIe® interface is a critical backbone for moving data at high bandwidth between the host processor and connected peripherals in modern enterprise and cloud data centers. The incumbent PCIe generation, PCIe 4.0, delivers 16 GT/s data transfer speed, achieving full duplex aggregate bandwidth of 64 GB/s in an x16 implementation. However, with the advent of cloud computing, faster networking, artificial intelligence (AI) and machine learning (ML), the demand for greater bandwidth has continued to accelerate and has pushed PCIe 4.0 to its limits. To keep pace with this unprecedented data-centric age, the industry is now moving to adopt the latest PCIe generation, PCIe 5.0. PCIe 5.0 scales up the data transfer rate to 32 GT/s, effectively doubling the data rate over the previous generation. The resultant 128GB/s full duplex bandwidth of an x16 PCIe 5.0 interface is sufficient to support up to 400 GbE links, future-proofing new hardware designs for upcoming data needs. The PCIe interface also enables NVMe SSDs – the new industry standard for blazing fast storage. Now, with the latest PCIe 5.0, NVMe SSDs are going to be even faster. In this blog post, AMD and Samsung collaborate to showcase the truly impressive potential of PCIe 5.0, which has the ability to transform and allow further advancement in high-performance applications. Next-generation Server Architecture – AMD EPYC™ 9004 Series Processors The AMD EPYC 9004 Series CPUs amplify the history of AMD x86 architecture innovations and record-breaking performance↗ with next generation 5nm technology. At the same time, these CPUs introduce support for high-performant DDR5 DIMMs and fast PCIe 5.0 I/O. AMD EPYC 9004 Series CPUs support 12 memory channels with 2 DIMMs/channel capability, delivering the resources needed for memory-hungry AI, ML, HPC, and large in-memory computations. The new AMD EPYC 9004 Series CPUs also uniquely provide 128 PCIe 5.0 lanes in a 1-socket server, and up to an astounding 160 PCIe 5.0 lanes in 2 socket servers. With high PCIe 5.0 lane counts, the 4th Generation AMD EPYC CPUs meet the performance demands of today’s AI and ML applications, as well as the increasing use of accelerators, GPUs, FPGAs and high-capacity LAN cards – natively delivering 4X the I/O throughput of competing solutions¹. Industry-leading PCIe 5.0 SSD – Samsung PM1743 Samsung’s PCIe Gen 5.0 PM1743 SSD has been designed specifically for use in servers and enterprise storage systems. With a proprietary controller, the PM1743 strengthens performance and reliability, providing enterprise users with expanded system capacity suitable for heavy workloads. The PM1743 offers a high random read speed ― a key metric for server storage performance ― tested at up to 2.5M IOPS. It delivers sequential read speed of up to 14GB/s, 2X faster than PCIe 4.0 SSD, along with 7.1 GB/s of sequential write and 360 KIOPs of random write performance. This drive comes in two form factors – U.2 and E3.S – with a wide range of capacities to suit the diverse needs of OEMs worldwide. I/O Performance from AMD EPYC 9004 Series CPU and Samsung PM1743 SSD One of the biggest challenges with storage use cases is that with a large number of SSDs, it’s possible to bottleneck on CPU and/or memory bandwidth at a relatively small number of SSDs. The real test is to see how well a storage system with PCIe Gen 5.0 SSDs can perform. To that end, as a global leader in NVMe SSDs, Samsung collaborated with AMD to demonstrate the performance of its latest AMD EPYC 9004 Series CPU for storage use cases. For this study, the proprietary AMD “Titanite” reference platform was used, along with Samsung’s PM1743 SSD. Titanite is equipped with an AMD 9654 96-core processor and is populated with 768 GB of DDR5 RDIMM. It supports 24 U.2 PCIe 5.0 x4 SSDs. Full details of the system configuration are shown in Table 1. Table 1:

_{*Disclaimer-Performance may vary on system conditions.} Flexible I/O Tester (FIO) is a widely used storage performance benchmarking tool. A series of FIO tests were performed using this tool to characterize the performance of the 4th Generation AMD EPYC processor under various storage workloads with all 24 slots of SSDs populated with 1.92TB PM1743. These tests show that the performance scales linearly as more drives are plugged in for testing, from one to 24 drives. For a fully populated server, we achieved > 220GB/s of sequential read, >63,000 MB/s for sequential write, >37 MIOPs of random read and > 14,000 kIOPs of random write performance. The graphs below show the raw throughput data of the FIO tests on the AMD “Titanite” reference platform with Samsung PM1743 SSD.

Conclusion The results achieved in this study are truly exceptional and demonstrate how AMD and Samsung can address specific performance challenges and provide solutions for applications with virtually insatiable demands for data. Learn more about Samsung’s PCIe Gen 5.0 PM1743 SSD and the new AMD EPYC 9004 Series Processor↗.

---

_{1. EPYC-036: One AMD EPYC 9004 CPU with 128 PCIe® 5 lanes has 4x the IO capacity of Intel Xeon Scalable Ice Lake CPUs with a max 64 PCIe4 lanes per https://ark.intel.com/↗. PCIe5 which operates at 32 GT/s and a 3.94 GB/s transfer rate where PCIe4 only operates at 16 GT/s and 1.97 GB/s transfer.}