Cost-Effective Memory for HPC, AI and Enterprise Storage Planning

With DRAM costs rising dramatically due to increasing AI-driven demand and constrained supply, fully populating memory in systems is no longer always economically justified. As memory becomes an increasingly significant component of system cost, organizations must carefully balance performance requirements against infrastructure budgets. The terms RAMageddon and RAMpocalypse have emerged to describe the unprecedented increase in DRAM pricing and the resulting pressure across the technology industry, from consumer devices to hyperscale data centers. In this paper, we examine practical strategies available to maximize workload performance while minimizing memory-related expenditure in this constrained market.

Using Lenovo ThinkSystem platforms, we present a series of price/performance studies spanning HPC, enterprise AI, and storage workloads, investigating where memory capacity can be reduced without materially impacting application performance or overall data center productivity. The workloads evaluated represent a broad range of academic and industrial use cases, including computer-aided engineering (CAE), computational fluid dynamics (CFD), molecular dynamics, computational chemistry, weather and climate modeling, media processing, AI training, AI inference, and enterprise storage.

By analyzing these workloads collectively, we identify memory population and sizing strategies that deliver the optimal balance of cost and performance for mixed-workload environments. The results provide practical guidance for organizations seeking to maximize infrastructure value, increase resource efficiency, and maintain user productivity during a period of sustained memory cost inflation.

1. Introduction
2. Understanding the performance of HPC Microbenchmarks with underpopulated memory channels   
3. Impact of MRDIMMs and Underpopulated Memory Channels on Real-World HPC Workloads   
4. Impact of Underpopulated Memory Channels on AI Training and Inference Workloads
5. Lenovo DSS-G Storage Solution with reduced memory
6. Conclusions and Recommendations
Appendix A1 – List of Codes in HPC Benchmark Suite

To view the document, click the Download PDF button.

Related product families

Product families related to this document are the following:

• Artificial Intelligence
• High Performance Computing
• MLPerf Benchmark
• Memory
• ThinkSystem SC750 V4 Server
• ThinkSystem SR675 V3 Server