Important lessons learned
A potent solution for AI and database infrastructures is provided by Micron DDR5 128GB memory and 5th Gen AMD EPYC CPUs with 128 cores. The computational complexity of huge model sizes and extensive datasets, which are features of AI, machine learning, data analytics, and IMDB applications, is successfully managed by this technique.
| SVM 1.3x Improves AI/ML support vector machine (SVM) performance by 1.3x.1.3x higher bandwidth use due to higher memory clock speeds on Micron 128GB RDIMMs, along with increased core counts on 5th Gen AMD EPYC processors. | IMDB REDIS 1.2x Improves IMDB Redis performance by 1.2x for a set:get ratio of 1:10.30% Delivers 30% better average latency and 60% better p99 latency. | SAP SD 201k users Improves SAP Sales and Distribution (SAP SD) benchmark. With 30% higher memory capacity and 30% higher clock, SAP SD scores higher than previous six-socket performance score. |
- Pairing Micron 128GB DDR5 RDIMMs with 5th Gen AMD EPYC processors (codenamed Turin).
- As compared to Micron 96GB DDR5 RDIMMs and 4th Gen AMD EPYC processors with 96 cores (codenamed Genoa).
High-capacity memory and substantial processing power are necessary for modern data centers to support enterprise machine learning (ML) and artificial intelligence (AI) programs by running a range of workloads. Micron DDR5 128GB RDIMMs and 5th Gen AMD EPYC processors work together to provide exceptional performance and capabilities for a variety of server workloads that data centers handle, such as hosting demanding corporate applications and powering expansive cloud-based infrastructures.
In this blog, Micron present the outcomes of benchmark tests for Redis in-memory database (IMDB), SAP SD, and AI/ML support vector machine (SVM), where Its hardware configuration contrasted the following:
- 5th Gen AMD EPYC processors (codenamed Turin) with Micron DDR5 128GB DIMMs
- With 4th Gen AMD EPYC processors (codenamed Genoa), Micron DDR5 96GB DIMMs
Its testing demonstrates that Micron DDR5 RDIMMs with increased capacity (128GB) and bandwidth (capable up to 8000 MT/s) improve performance for SVM, SAP SD, and Redis IMDB.
Hardware and system setup
The table below displays the specifics of the system architecture. Two systems, designated A and B in this blog, were compared. System A has Micron 96GB DDR5 DIMMs with 4th Gen AMD EPYC CPUs (96 cores), while System B had 128GB and 5th Gen. Both systems support a 12GB/core configuration; the 96-core CPU has 96GB across 12 memory channels, while the 128-core CPU has 128GB.
System A | System B | |
|---|---|---|
| Hardware | 4th Gen AMD EPYC processors (codenamed Genoa) | 5th Gen AMD EPYC processors (codenamed Turin) |
| Memory | Micron 96GB DDR5 4800 MT/sDual rank, 12 channels | Micron 128GB DDR5 6400 MT/sDual rank, 12 channels |
| CPU | Dual-socket AMD EPYC 9654 (96-core) | Dual-socket AMD EPYC (128-core) |
| Storage (for SVM) | Micron 9400 8TB (3) | Micron 9400 8TB (3) |
Support Vector Machine in AI/ML
SVM is a kind of machine learning technique that is frequently used to prepare datasets for a variety of cloud-based data science applications. Its processed a 2TB dataset using the SparkML engine and the Intel Hi-Bench architecture in tests.
Faster execution time
System B outperformed system A by 30% in terms of SVM execution time. This is mostly because system B’s processor has more cores, the 128GB memory modules offer more capacity and bandwidth, and the bandwidth is used efficiently.
Higher bandwidth use
Because of the faster memory rate (6400 MT/s vs. 4800 MT/s) and the extra Zen5 cores made possible by the 5th Gen AMD EPYC processors with 128 cores, data indicate that SVM on system B uses 1.3 times more bandwidth than system A.
The SVM can keep more data in memory because to system B’s larger capacity (128GB vs. 96GB), which reduces storage input/output. For both setups, maintained a consistent memory capacity of 12GB per core. Compared to the baseline configuration (system A), this method allowed us to separate the effects of more computation capacity and faster clock speed (memory).
Redis
Applications that need minimal latency can store and retrieve data using Redis, a fast in-memory database (IMBD). Memtier simulates a multithreaded and multiclient execution model by benchmarking Redis with multiple set:get operations.
Running Redis on system B (128GB and 128 cores) results in a 1.2x speedup. Additionally, the identical combination reduces p99 latency by 60% and average latency by 30%. Higher core counts, such as the 5th Gen’s 128 cores, can better utilize the Micron DDR5 128GB DIMMs’ greater capacity and bandwidth than earlier AMD EPYC CPU generations. Enterprise data centers can easily serve more users with the increased throughput made possible by the additional cores.
SAP Sales and Distribution (SAP SD)
An enterprise resource planning (ERP) software suite that is frequently used is called Systems Application and Products (SAP). As a component of the SAP ecosystem, it is composed of several smaller parts. Setting a new performance world record of 201,000 users for the SAP SD benchmark on a two-socket system, the component that includes all of the operations and processes for SAP Sales and Distribution (SAP SD) was benchmarked against the Dell PowerEdge R6725 server and outfitted with Micron DDR5 128GB RDIMMs and 5th Gen AMD EPYC processors.
That surpasses the top six-socket score. The increased number of benchmark users suggests that using Micron memory in conjunction with 5th Gen AMD EPYC CPUs on Dell PowerEdge servers for database use cases offers a performance advantage.
AI in data centers
For data center infrastructures to efficiently manage the computational complexity, massive model sizes, and extensive datasets typical of AI, machine learning, data analytics, and IMDB applications, high-capacity memory, high memory bandwidth, and low latency are essential. Micron workload findings demonstrate that Micron DDR5 128GB memory modules combined with 5th Gen AMD EPYC processors provide a potent solution for these situations.
