Quantum’s Myriad OS now supports Nvidia GPUDirect file access for fast GPU Server data feeding.
Myriad, announced in April 2023, is a containerized, unified, scale-out file and object storage software stack, based on key/value storage technology, for all-flash NVMe media. It’s designed to manage trillions of files and objects. In April this year Quantum added Tuxera Fusion File Share IP for server message block (SMB) file access which improves file service performance and stability with features like SMB Direct RDMA, SMB scale-out, SMB Multichannel, SMB witness protocol, and persistent file handles. GPUDirect enables direct data loading into GPU server memory using RDMA from remote NVME drives.
This latest Myriad development adds a parallel file system client, designed to fully enable GPUDirect storage capability, and offering on-the-fly client node deployments with cross platform compatibility – including Nvidia Grace Hopper with Grace Arm-based architectures. It provides an approach to building AI/ML (Artificial Intelligence/Machine Learning) infrastructure enabling customers to add graphics processing unit (GPU) nodes to Myriad clusters.
Jeff Mulder, Quantum’s chief development officer, stated: “The new client brings unique capabilities by running as a fully integrated GPUDirect Node on client systems. Unlike traditional parallel file system clients, which rely on other appliances to perform these operations on their behalf, Myriad’s client performs its own metadata operations, data reduction, and data protection operations. This approach minimizes common bottlenecks, lets customers maximize their GPU investment across architectures, and allows performance to scale with the number of parallel clients accessing the Myriad system.”
Myriad’s parallel client design takes a new approach to building out AI/ML infrastructure by installing directly on customer servers or workstations equipped with GPU cards, transforming their host workstation into operational Myriad GPU Nodes.
The new client is optimized for GPU-intensive workloads such as AI/ML model training and inferencing, high-performance computing (HPC) visualization and modeling, and video rendering. It’s designed to maximize GPU utilization and performance.
Multiple client nodes can be added to a Myriad cluster at any time, and are a new connection option alongside SMB, network file system (NFS), and planned S3 access that can be added to any Myriad share point on-demand.
The added GPUDirect support means that Myriad systems can compete for AI training and inference workloads with existing GPUDirect-supporting suppliers – such as DDN, IBM, Huawei, NetApp, Pure Storage, VAST Data and WEKA. It won’t be frozen out of new storage bids needed for AI/ML workloads featuring Nvidia GPU servers.
Lance Hukill, chief commercial officer for CHESA, an integrator for media and entertainment customers, said: “We are extremely impressed with the performance and capabilities of Myriad after testing it in our labs.”
The new Myriad parallel file system client will be made available for evaluation through Quantum’s Early Access Program. Existing Myriad customers are eligible to participate in the Early Access Program for the parallel file system client upon its release in Q1 2025, with general availability in the second half of 2025. Find out more about the Early Access Program here.
Quantum says its network of system integrators is ready to help integrate Myriad’s capabilities for content production needs such as animation/visual effects (VFX) rendering and AI/ML actions for rich media and video at scale.
