Ask a Server SSD Expert
Planning the right solution requires an understanding of your project's storage goals. Let Kingston's experts guide you.
Training a machine‑learning model can consume more CPU cycles, memory and I/O bandwidth than any other software task. A single state‑of‑the‑art model may read and write hundreds of terabytes of data across multiple epochs, pushing the limits of even the fastest storage arrays. With the emergence of the AI race, as every industry and government agency competes to push the envelope of AI performance, the need for a reliable, repeatable way to benchmark specific hardware resources for ML workloads has become essential.
The MLCommons group has pioneered this effort with MLPerf Storage v2.0, introducing innovative methods to evaluate storage performance in machine learning training and inference. This whitepaper explores how Kingston’s DC3000ME PCIe Gen5 NVMe U.2 SSDs perform under MLPerf’s benchmarking framework without requiring expensive GPU hardware. Using sleep emulation, MLPerf simulates real-world AI workloads, allowing organisations to test storage performance in a cost-effective and scalable way.
Why this matters? Fast storage keeps GPUs productive by feeding data quickly, minimising idle time and maximising compute efficiency. Kingston’s DC3000ME drives help organisations accelerate training cycles, reduce infrastructure costs and achieve scalable performance with fewer drives, making them a smart investment for modern AI deployments.
In this whitepaper, we:
- Examine the MLPerf v2 storage novel sleep emulation test method
- Introduce the different training and checkpointing workloads covered by MLPerf v2 storage
- Discuss the results of Kingston DC3000ME’s submission in the training and checkpointing workloads in PCIe Gen5 and Gen4 platforms
The results show that PCIe Gen5 NVMe solutions like DC3000ME present a compelling solution for organisations looking to achieve the highest accelerator utilisation and GPU active hours in their machine learning workflows. We dive deep into tunables, test methodology and post-mortem tuning and their impact on different workloads.
Kingston’s DC3000ME PCIe Gen5 NVMe U.2 SSDs are a high-impact solution for AI workloads. Whether you're training deep learning models or fine-tuning large language models, these drives deliver the speed, reliability and efficiency needed to stay ahead in the AI race.
Download whitepaper
