Unlock a New Tier of Ultra-Fast Storage to Accelerate HPC Workloads

Turn Siloed GPU Server Local Storage Into a New Tier 0 for Simulations and Model Checkpointing

The investment to build infrastructure for high-performance computing is substantial – large GPU computing clusters, fast networks, and massive storage systems to feed data to the GPUs. This all adds up to cost constraints, power constraints, and valuable GPU assets that sit idle all too often.

• Legacy parallel file systems are complex to manage, with limited support and an uncertain roadmap
• Even modern NAS systems can not provide the performance to keep GPUs utilized
• Power constraints are a major concern for HPC infrastructure teams

Meanwhile, GPU servers have fast local storage that typically goes unused because it is not usable by all of the GPUs in the cluster, it is not protected, and difficult for data users to manually manage the data.

You may have 100s of Terabytes – even many Petabytes – of fast NVMe storage capacity just sitting in your GPU servers already – unused and stranded. You can start to use that capacity in minutes with Hammerspace, so you can accelerate your data pipelines, complete simulations faster, and reduce checkpointing times.

Hammerspace Set a New Record in MLPerf 1.0 Benchmark with Tier 0 Storage

The MLCommons MLPerf Storage benchmark is intended to demonstrate the performance of various storage systems for simulated machine learning workloads, so that technical buyers and decision makers have some criteria when evaluating storage system performance for machine learning, deep learning, and other forms of GPU computing.

By eliminating network constraints, scaling performance linearly and delivering unparalleled financial benefits, Tier 0 sets a new standard for AI and HPC workloads.

Best of all, Tier 0 achieves linear performance scaling by processing data directly on GPU-local storage, bypassing traditional bottlenecks. Using Hammerspace’s data orchestration, Tier 0 delivers data to local NVMe drives, protects it and seamlessly offloads checkpointing and computation results.