5th Gen AMD EPYC CPUs : The Engine Behind Google Cloud’s Next Leap

Industry response to our 5th Gen AMD EPYC™ processors introduced in late 2024, has been phenomenal. We’re deeply grateful to our partners and customers for their continued support. As a result, AMD EPYC now commands over 40% of server CPU market share¹—a milestone driven by exceptional performance, energy efficiency, low total cost of ownership, and strong collaboration with cloud and enterprise ecosystem partners.

In my previous blog posts, I explored how our latest generation of processors accelerate performance across enterprise, HPC, and AI workloads. If you missed them, you could find those insights here:

• 5th Gen AMD EPYC Processors Lead Enterprise and Cloud Workloads Forward
• 5th Gen AMD EPYC Processors Elevate HPC and AI Workloads to New Heights

The new 5th Gen EPYC processors deliver significant generational improvements—up to doubling performance in many scenarios. Powered by the new “Zen 5” cores, they offer up to 17% better single-threaded performance on enterprise and cloud workloads² and scale up to 192 cores per processor, a 50% increase in maximum core count over the previous generation. Memory bandwidth has also advanced, supporting speeds up to 6400 MT/s and up to 8 TB per socket, enabling seamless handling of large, complex workloads. Confidential Computing enhancements now include support for Trusted I/O, allowing devices enabled with the PCI TEE Device Interface Security Protocol (TDISP) to be securely bound to a confidential guest virtual machine.

EPYC in Google Cloud

Since 2019, AMD has collaborated with Google Cloud to expand support across general-purpose, high-performance, and confidential computing instances. In 2025, the collaboration reached a new milestone with the launch of C4D and N4D virtual machines powered by 5th Gen AMD EPYC processors, delivering faster CPU clock speeds, enhanced memory bandwidth, and robust security features. At launch, Google Cloud also introduced its first AMD processor-based bare metal instances, and made Confidential VMs available simultaneously, offering additional options to impact performance and protection. C4D instances support up to 384 vCPUs and up to 3 TB of memory, instances with local SSDs, and advanced networking and storage options — providing customers with flexible, high-performance solutions for demanding workloads. N4D instances support up to 96 vCPUs and 768 GB of DDR5 memory. Google positions C4D as its consistent high performance general-purpose instance series, with premium features such tier 1 networking and the Confidential option, while N4D provides cost efficiency for less performance-sensitive workloads and a custom VM option. Google Cloud made C4D generally available in June 2025 and brought N4D to GA in November of this year.

Workload Performance Characterization 

In this blog, I’ll share both generational and competitive performance insights for AMD 5th Gen EPYC CPU-based C4D and N4D instances on Google Cloud. The benchmarks focus on 16 vCPU instance sizes — one of the most widely used VM configurations for general-purpose computing. We evaluated a range of workloads, including server-side Java®, relational databases, web serving, in-memory analytics and media processing.

To provide meaningful comparisons, we measured how C4D and N4D perform relative to the previous-generation C3D instances, as well as against C4 instances powered by Intel’s 5th Gen Xeon® processors. The analysis includes both raw workload performance and performance-per-dollar metrics on 16 vCPU standard (4:1 GB to vCPU, no local SSD) instances.

1. General Purpose Computing

The SPEC CPU® 2017 benchmark is a highly regarded, regulated, and audited industry standard for assessing compute-intensive workloads. It rigorously tests processors, memory subsystems, and compilers across a variety of systems, providing valuable insights into performance and efficiency. This benchmark serves as a critical tool for evaluating the capabilities of modern computing architectures. This blog focuses on estimated SPECrate® 2017 Integer performance.

As shown in Figure 1, C4D instances deliver approximately 1.27x the estimated SPECrate® 2017 Integer performance, and N4D instances deliver 1.29x the estimated performance compared to the prior generation C3D. C4D instances deliver an estimated 1.20x and N4D instances deliver an estimated 1.23x the integer performance compared to C4 standard instances based on Intel 5th Gen Xeon architecture. When evaluating performance-per-dollar, C4D demonstrates around 1.29x (est.) and N4D demonstrates around 1.43x (est.) the efficiency of C4.

Figure 1: Estimated SPECrate® 2017 Integer performance^3-6

2. Server-side Java

Java is a widely adopted programming language, renowned for its portability and versatility, which allows developers to build applications that run reliably across a broad range of platforms. For this analysis, we used a server-side benchmark — a well-established tool for evaluating Java application performance in a simulated enterprise IT environment. This benchmark models workloads such as point-of-sale transactions, online commerce, and data-mining operations, making it highly relevant to JVM vendors, hardware manufacturers, developers, and researchers.

One of the key metrics from server-side benchmarking is max-jOPS (Java operations per second), which measures the maximum throughput of Java-based e-commerce transactions and reflects the system’s peak processing capacity. As shown in Figure 2, C4D instances deliver approximately 1.24x and N4D instances deliver approximately 1.18x the Java multi-instance max-jOPS compared to C3D. C4D instances deliver approximately 1.22x and N4D instances deliver approximately 1.16x the Java multi-instance max-jOPS compared to C4 instances based on Intel 5th Gen Xeon architecture. When evaluating performance-per-dollar, C4D demonstrates around 1.31x and N4D demonstrates around 1.36x the efficiency of C4.

Figure 2: Server-side Java® multi-instance max-jOPS^7-10

3. Relational Databases

MySQL™ is one of the most widely used open-source database management systems, trusted globally across enterprise and cloud-native environments for both decision support and transactional workloads. For this evaluation, we used the TPROC-C benchmark, which is derived from the industry-standard TPC-C™, designed to measure performance in transaction processing systems.¹¹

Figure 3 presents the results of this benchmark, showing that C4D instances deliver approximately 1.57x and N4D instances deliver approximately 1.53x the MySQL TPROC-C throughput compared to C3D. C4D instances deliver approximately 1.53x and N4D instances deliver approximately 1.48x the MySQL TPROC-C throughput compared to C4. When looking at performance-per-dollar, C4D achieves around 1.63x and N4D achieves around 1.73x the efficiency of C4.

Figure 3: MySQL™ TPROC-C Transactions Per Minute^7-10 

4. Web Serving 

NGINX™ is a widely used web server known for its flexibility and efficiency in handling client requests. It can function as a standalone server or enhance performance and security by acting as a reverse proxy, load balancer, mail proxy, or HTTP cache. To evaluate its performance, we used the popular WRK benchmarking tool, which generates substantial HTTP traffic to simulate real-world web serving conditions.

Figure 4 shows that C4D instances deliver approximately 1.83x and N4D instances deliver approximately 1.70x the NGINX throughput of C3D. C4D instances deliver approximately 1.75x and N4D instances deliver approximately 1.62x the NGINX throughput of C4. In terms of performance-per-dollar, C4D achieves around 1.87x and N4D achieves around 1.89x the efficiency of C4.

Figure 4: NGINX Benchmark using WRK^7-10

5. In-Memory Analytics

Redis™ is a widely used in-memory data store that functions as a key-value database, cache, and message broker, with optional durability features. Its high performance and support for diverse data types—such as strings, lists, and maps — make it ideal for cloud-native environments and use cases like streaming, microservices, and real-time analytics.

To evaluate Redis performance, we used the redis-benchmark tool, which simulates multiple clients connecting to the server and measures the average number of requests handled per second. This provides a clear view of how well Redis performs under load. The results are an average of Get and Set.

Figure 5 shows that C4D instances deliver approximately 1.32x and N4D instances deliver approximately 1.82x the Redis throughput of C3D. C4D instances deliver approximately 1.63x and N4D instances deliver approximately 1.52x the Redis throughput of C4. In terms of performance-per-dollar, C4D achieves around 1.74x and N4D achieves around 1.77x the efficiency of C4.

Figure 5: In-Memory Analytics Redis™ redis-Benchmark, Get and Set Average^7-10

6. Media Processing

Media processing performance is critical for a wide range of workloads — from video encoding and transcoding to collaboration platforms and streaming services. FFmpeg is a powerful, open-source multimedia framework that includes a rich set of libraries, codecs, and tools for handling video, audio, and other media formats. Known for its command-line efficiency, FFmpeg is widely used for tasks such as encoding, transcoding, editing, scaling, post-production, and ensuring standards compliance.

For this evaluation, we averaged two encode and two transcode benchmarks using VP9 and h.264 codecs. These tests highlights C4D’s ability to manage high-resolution video content efficiently.

Figure 6 shows that C4D instances deliver approximately 1.47x and N4D deliver approximately 1.44x the FFmpeg throughput of C3D. C4D instances deliver approximately 1.70x and N4D delivers approximately 1.66x the FFmpeg throughput of C4. In terms of performance-per-dollar, C4D achieves around 1.81x and N4D achieves around 1.93x the efficiency of C4.

Figure 6: FFmpeg encode and transcode performance^7-10

Conclusion:

Media processing performance is critical for a wide range of workloads — from video encoding and transcoding to collaboration platforms and streaming services. FFmpeg is a powerful, open-source multimedia framework that includes a rich set of libraries, codecs, and tools for handling video, audio, and other media formats. Known for its command-line efficiency, FFmpeg is widely used for tasks such as encoding, transcoding, editing, scaling, post-production, and ensuring standards compliance.

For this evaluation, we averaged two encode and two transcode benchmarks using VP9 and h.264 codecs. These tests highlights C4D’s ability to manage high-resolution video content efficiently.

Figure 6 shows that C4D instances deliver approximately 1.47x and N4D deliver approximately 1.44x the FFmpeg throughput of C3D. C4D instances deliver approximately 1.70x and N4D delivers approximately 1.66x the FFmpeg throughput of C4. In terms of performance-per-dollar, C4D achieves around 1.81x and N4D achieves around 1.93x the efficiency of C4.