Powering AI, HPC & Enterprise with 5th Gen AMD EPYC processors: A Game Changer for Modern Workloads

At Boston, we’ve spent over 30 years partnering with industry leaders like AMD and Supermicro to deliver cutting-edge server solutions. Today, we’re thrilled to announce the integration of 5th Generation AMD EPYC™ processors also called EPYC 9005 series, previously codenamed “EPYC Turin” into our Boston Labs. These processors, built on the new Zen 5 and Zen 5c core architectures, are designed to tackle the most demanding workloads in enterprise, AI, and high-performance computing. Let’s dive into what makes these processors revolutionary and how they can transform your data centre.

Picture: AMD EPYC 9755 CPU

Why the 5th Generation AMD EPYC Processors Matters?

• Up to 192 cores and 384 threads per socket 
• Up to 6TB of DDR5-6400MT in 12 memory channels per socket
• 128 PCIe 5.0 lanes in single socket and up to 160 lanes in dual socket configuration 
• Up to 384MB of L3 cache
• CLX 2.0 support
• Built-in security for VM encryption

AMD EPYC 9005 Series Processors Specification:

Decoding EPYC 9005 Core Types and SKU Suffixes

AMD’s 5th Gen EPYC processors feature two core types to optimise performance and efficiency: the Zen 5 core, designed for single-threaded, latency-sensitive workloads like enterprise databases, offers up to 128 cores with 8 cores per CCX and 32MB shared L3 cache, supporting up to 16 CCDs in the SP5 socket. In contrast, the Zen 5c core, tailored for high-density, multi-core applications such as AI training and HPC simulations, provides up to 192 cores with 16 cores per CCX, also with 32MB shared L3 cache, and up to 12 CCDs in the SP5 socket, leveraging a compact layout for maximum performance per watt.

You might have noticed that some CPU SKUs have "P" and "F" suffixes. The AMD EPYC processors with a "P" suffix, such as the EPYC 9655P, are designed for single-socket configurations only, making them cost-effective and efficient for workloads that don't require dual-socket setups. These SKUs offer the same performance and features as their "non-P" counterparts, but they come with a much lower price tag. In contrast, "F" suffix SKUs, like the EPYC 9575F, are frequency-optimised for highest per-core performance, targeting workloads such as GPU-powered AI solutions that benefit from higher clock speeds, with boosts up to 5GHz. These suffixes help tailor EPYC processors to specific use cases, balancing core count, frequency, and power efficiency.

System-on-Chip (SoC) Design: Simplicity and Security

AMD EPYC processors utilise a System-on-Chip (SoC) design that integrates CPU cores, memory controllers, and security features into a single package, simplifying server design by reducing reliance on external chipset components for improved efficiency and reliability. The embedded AMD Secure Processor enhances security through features like a secure root of trust, Transparent Secure Memory Encryption (TSME), and Secure Encrypted Virtualisation (SEV). Together, these capabilities ensure robust protection for sensitive workloads across cloud, enterprise, and edge environments.

AMD Infinity Fabric: High-Speed Connectivity

The AMD Infinity Fabric™  is the backbone of 5th Gen EPYC internal connectivity. It uses sixteen 36Gb/s Infinity Fabric links to connect the I/O die with each CPU die. In memory-speed-optimiSed EPYC 9005 Series processors, two links connect to each CPU die, delivering up to 72Gb/s of connectivity. This ensures low-latency, high-bandwidth communication for demanding workloads.

Compatibility

• Supermicro H13 generation motherboards are able to support EPYC 9005 processors after a simple BIOS update but only the CPU SKU’s with the TDP of up to 400W
• The latest Supermicro H14 motherboards can support the entire EPYC 9005 line-up out-of-the box, and are also compatible with the previous 4th generation of EPYC CPU’s 

Workload Performance

Benchmark testing at Boston Labs

For our benchmark analysis, we evaluated the EPYC 9755 CPU and compared it directly to its 4th-generation EPYC counterparts. We tested CPUs with 128 cores and 96 cores for a head-to-head comparison. All CPUs were installed in a dual-socket server configuration.

• Supermicro 2U H14 Hyper Server (dual socket)
• 16x 64GB DDR5-4800MT memory
• Rocky Linux 9.4

This benchmark assesses data compression performance, rated in MIPS (millions of instructions per second). As the graphs demonstrate, the EPYC 9755 provides enhanced performance in these workloads.

We observed significantly reduced rendering times in Blender scenes, demonstrating excellent performance with approximately a threefold reduction in rendering time.

In the C-Ray test - a simple multi-threaded ray tracer designed to evaluate floating-point CPU performance. The results also showed a substantial performance improvement in ray tracing.

In the PostgreSQL 15 Benchmark test, the results highlight a significant performance improvement in Database Read-Only (TPS) performance.

Based on the test results from our lab, the 5th-generation EPYC CPU performs exceptionally well across a wide range of application performance benchmarks.

Testing in Boston Labs: See the Power in Action!

At Boston, we’re not just talking about specs—we’re putting these processors to the test. Our Supermicro servers, fully validated for AMD EPYC 9005 series, are now available for testing in our labs. We invite you to explore how these processors can redefine modern computing for your workloads.

Off-the-shelf solutions won’t cut it for cutting-edge challenges in AI and HPC. At Boston, we understand that every workload is unique. That’s why we’re here to help you unlock the full potential of 5th Gen AMD EPYC CPUs and Supermicro’s advanced systems.

Ready to experience the power of 5th Gen EPYC? Request a test drive by emailing [email protected] or calling +441727 876100. One of our experienced sales engineers will guide you through building the perfect solution for your needs.

Author

Alvin Leung

Field Application Engineer

Boston Limited

Tags: amd, epyc.9005 series, cpu, amd epyc, amd 5th gen, ai, hpc, workloads