New MinIO AIStor integrations use NVIDIA infrastructure to enhance AI storage at a multi-exabyte scale.
MinIO announced three upcoming advancements to MinIO AIStor platform at NVIDIA GTC, aimed at enhancing support for AI workloads within the NVIDIA AI ecosystem. The new integrations aim to improve AI infrastructure efficiency and simplify management, allowing personnel to focus on strategic tasks.
The new MinIO AIStor features include:
- Support for NVIDIA GPUDirect Storage (GDS) for object storage:Â Improves CPU efficiency on the NVIDIA GPU server by bypassing the traditional data path through the CPU, allowing more compute resources for AI data processing and reducing infrastructure costs with support for Ethernet networking fabrics.
- Native integration with NVIDIA BlueField-3 networking platform:Â Reduces object storage total cost of ownership (TCO) while maintaining high performance and supporting data-driven and AI workloads at a large scale for enterprise environments.
- Incorporation of NVIDIA NIM microservices into AIStor promptObject inference:Â Simplifies deployment and management of inference infrastructure while introducing AIStor’s new S3 API promptObject, which enables interaction with unstructured objects similar to engaging with an LLM, improving inference speed through model optimizations for NVIDIA hardware.
Maximizing the utilization and efficiency of AI compute infrastructure (GPUs and CPUs)
NVIDIA GPUDirect Storage (GDS) was initially designed to work with InfiniBand, requiring specific hardware. NVIDIA has extended the benefits of GDS and InfiniBand to Ethernet networks. This innovation provides flexibility, scalability, and cost efficiency, making it a solution for accelerated AI adoption at scale for Enterprises.
MinIO AIStor is designed to maximize per-node network bandwidth to efficiently supply data to GPUs. MinIO AIStor’s GDS for object storage implementation leverages Ethernet fabrics by establishing a direct data path between MinIO AIStor and NVIDIA GPU memory. This improves overall GPU server efficiency, increasing resources for additional AI-related compute. MinIO AIStor and NVIDIA GDS deliver an efficient, adaptable solution for scaling AI infrastructure, and create a streamlined, fast pipeline that turns data lakes into high-speed AI/ML training environments.
Additionally, MinIO AIStor becomes the first and only object storage software to run natively on NVIDIA’s BlueField-3 Data Processing Unit (DPU), made possible by AIStor’s remarkably compact ~100MB footprint. This ultra-efficient, low-cost architecture completely eliminates the need for separate x64 CPUs, transforming what were already commodity storage servers into MinIO and NIC-powered JBOFs (Just a Bunch of Flash).
MinIO AIStor leverages Arm’s Scalable Vector Extension (SVE) instruction set to enhance object storage performance and inline data management directly on NVIDIA BlueField-3 DPUs. This integration enables compatibility with NVIDIA Spectrum-X networking and GPUDirect Storage, reducing data movement overhead and improving GPU server efficiency for AI and high-performance workloads.
Maximizing inference performance and streamlining infrastructure management
MinIO AIStor simplifies AI-powered interactions with stored objects through the incorporation of AIStore promptObject into NVIDIA NIM microservices inference infrastructure. NIM provides pre-built Docker containers and Helm charts, and GPU Operator, which automates the deployment and management of drivers and the rest of the inference stack on the NVIDIA GPU server.
MinIO AIStor, leveraging NVIDIA NIM microservices, accelerates time to value and frees personnel from manual data pipeline and infrastructure building, enabling them to concentrate on strategic AI initiatives. In addition, NVIDIA NIM model optimizations for NVIDIA hardware deliver accelerated promptObject inference results.
These new features and integrations are open to beta customers under private preview. MinIO AIStor support for NVIDIA GDS and native integration with NVIDIA BlueField-3 networking platform will be released in alignment with NVIDIA’s GA calendar.
For more information, visit min.io.
