Combine HPC performance with standards-based NAS simplicity to accelerate data pipelines and maximize GPU resource utilization.
What is Hyperscale Nas?
Traditional storage options don’t meet the needs of enterprise AI and GPU computing. Scale-out NAS systems can not meet the performance or scale requirements, while legacy HPC parallel file systems are too complex and lack enterprise features.
Hyperscale NAS is a new NAS architecture – a standards-based parallel file system architecture – that combines the performance and scale of HPC file systems with the simplicity of enterprise NAS, so you can meet the extreme performance requirements of GPU computing while reducing costs and complexity.
The world has entered a new era of enterprise AI and deep learning. Organizations in every industry are training generative AI models in order to gain efficiencies, accelerate innovation, and drive their business forward. Training these models requires access to large data sets, distributed GPU resources, and extremely fast storage to feed large GPU clusters.
Read More
Keeping these GPUs fully utilized requires extreme parallel performance that is not possible with scale-out NAS, and has previously only been possible with complex HPC parallel file systems that are difficult to deploy and manage and lack enterprise features.
Like other parallel file systems, Hyperscale NAS delivers high-throughput, low-latency performance to compute clients, and that performance scales linearly to thousands of nodes.
Visit the Parallel NFS Resource CenterBut unlike other legacy parallel file systems that rely on complex proprietary clients to establish that performance, Hyperscale NAS uses capabilities built into the NFSv4.2 client that are now part of every modern Linux OS. These capabilities, specifically the use of Parallel NFS with Flex Files, were engineered by Hammerspace and contributed to the Linux specification years ago to create an intelligent parallel file system client built right into Linux.
Deliver high-throughput, low-latency performance to keep large GPU and CPU clusters utilized and minimize idle time.
Reduce server and networking costs by up to 2x, while increasing performance with an efficient direct data path architecture.
No proprietary client, no exotic networking technology, and no special storage technology required.
Configure clusters for maximum reliability, availability, and fault tolerance. Policy-based data protection is efficient and automated.
Summary The MLCommons MLPerf Storage benchmark is designed to demonstrate the performance of data storage platforms supporting AI/ML workloads. The aim is to provide AI system architects and decision makers…
This year, Hammerspace submitted results for the MLPerf v1.0 Storage Benchmark for the first time. This technical brief summarizes the results of that benchmark test
Hyperscale NAS is built for the next data cycle powered by AI and machine-generated data. It’s the first architecture to merge the simplicity of enterprise NAS with the extreme performance…
Organizations in every industry are grappling with how to realize the promise that artificial intelligence and deep learning (AI/DL) technologies can extract previously hidden value from unstructured data to improve…
Hammerspace CEO David Flynn discusses the intersection of HPC and Enterprise to power AI model training, and introduces the new Hyperscale NAS technology being adopted worldwide to power GPU computing.
