Micron and AMD deliver exceptional performance

Micron and AMD deliver exceptional performance for cloud-native workloads with 96GB DDR5 and 4th Gen AMD EPYC™ processors

Micron recently announced the availability of high-performance RDIMM solutions to help address computationally intensive artificial intelligence (AI), data analytics and memory-focused workloads. Working in collaboration with AMD, our joint goal was to elevate high-performance computing (HPC) workloads by harnessing the capabilities of Micron DDR5 and the advanced features of 4th Gen AMD EPYC™ processors. Since then, both companies have achieved remarkable progress, including the successful validation of new capacities like 24GB, 48GB and 96GB DDR5 DIMMs in January 2023. This blog post focuses on showcasing the impressive performance of the new 96GB DDR5 in conjunction with 4th Gen AMD EPYC processors.

Capitalizing on strengths to gain improvements

By collaborating with AMD, Micron capitalized on these latest AMD EPYC processors' cloud-native computing strengths, offering superior power efficiency. These improvements target sustainability goals and, along with high performance per watt, aligns perfectly with the key metrics widely used in the data center industry.

Here are the key strengths of this combination:

• Leading-edge performance: AMD EPYC 9754 processors are designed to meet the demands of cloud-native workloads. With up to 128 physical cores per processor and generous L3 cache sizes (up to 384MB per processor), they provide a high level of parallel processing power. This power enables efficient execution of concurrent tasks and supports the scalability required by cloud-native applications. 
• Impressive DDR5 speeds: Micron DDR5 memory modules are designed for remarkable speeds of up to 51.2 GB/s, ensuring fast data access and transfer within the system. This high bandwidth allows for seamless handling of large datasets and supports the rapid data processing required by cloud-native workloads.
• Cutting-edge processing: Micron's advanced 1β (1-beta) node processing technology brings several benefits to the table. It offers a 15% improvement in power efficiency, enabling more compute power while minimizing energy consumption. Additionally, there is a 35% increase in bit density with a 16Gb per die capacity than in the previous-generation 1α (1-alpha), allowing for higher memory capacities and improved overall system performance.
• Enhanced data integrity and reliability: The integrated error correction code (ECC) parity in Micron DDR5 memory ensures data integrity by detecting and correcting memory errors. This feature is crucial for cloud-native workloads that handle large amounts of critical data, providing an added layer of protection against potential data corruption. The presence of ECC parity enhances the overall reliability and stability of the system.
• Energy efficiency and performance: The latest 128-core processor from AMD focuses on energy efficiency, offering exceptional power efficiencies while supporting cloud-native workloads. The processor boasts proven RAS (Reliability, Availability, and Serviceability) capabilities and broad x86 hardware and software compatibility. Our testing reveals an outstanding performance/watt improvement of 2.68x compared to the previous generation.

By leveraging the power of AMD EPYC 9754 processors, the high-speed and efficient Micron DDR5 memory, and the robust ECC parity feature, we see an optimal solution for cloud-native workloads. This combination enables high-performance computing, efficient data processing, broad memory capacities, and reliable operation, all of which are essential for cloud-native applications in modern data center environments. 

Configuring and benchmarking cloud in-memory data stores

To simulate a workload that closely resembles Micron’s own IT cloud-native environment, we selected the Redis YCSB Proofpoint Workload D. This workload encompasses 250 million rows, each with a record size of 2KB, resulting in a total database size of 925GB. 

Testing setup involved running 64 instances with one Redis server and four clients, with a focus on performance and scaling. Performance was measured using operations per second (ops/s), and we scaled the workloads while ensuring that the latency remained the same or lower than in the previous generation.