At the NVIDIA GTC 2025 event held in San Jose, California, MinIO, Inc. introduced several enhancements to its MinIO AIStor platform, specifically designed to strengthen its integration within the NVIDIA AI ecosystem. These upgrades aim to optimize the performance and usability of AI infrastructures, allowing organizations to streamline management and redirect efforts toward more strategic AI initiatives.
Key Features of MinIO AIStor
- NVIDIA GPUDirect Storage (GDS) Support: This feature increases CPU efficiency on NVIDIA GPU servers by bypassing the traditional data path through the CPU, thus enhancing compute resources for AI data processing and reducing costs by leveraging Ethernet networking.
- Integration with NVIDIA BlueField-3 Networking: This native support lowers total cost of ownership (TCO) for object storage while offering top-tier performance. It optimizes data-driven workloads on a massive scale, catering to modern enterprise requirements.
- NVIDIA NIM Microservices Incorporation: The integration simplifies deployment and management of inference infrastructure. It introduces the new S3 API promptObject within AIStor, allowing easier interaction with unstructured data, akin to engagements with large language models (LLMs), and enhances inference speed through model optimizations for NVIDIA hardware.
Enhancing Efficiency and Scalability
NVIDIA previously required specialized hardware, such as InfiniBand, for GDS. Now, the extension to Ethernet networks brings better flexibility and cost-effectiveness, making it easier for enterprises to adopt accelerated AI solutions at scale. AIStor is designed to fully utilize the per-node network bandwidth, creating a direct data path between the storage and NVIDIA GPU memory. This capability significantly boosts the efficiency of GPU servers and enhances resources available for AI computations.
The collaboration between MinIO AIStor and NVIDIA GPUDirect Storage is set to create an efficient and adaptable AI infrastructure for enterprises, enabling them to convert data storage systems into dynamic environments suitable for high-speed AI and machine learning training.
MinIO AIStor and NVIDIA BlueField-3 DPU
MinIO AIStor stands out as the first object storage solution to operate natively on NVIDIA’s BlueField-3 Data Processing Unit (DPU). This is facilitated by AIStor’s compact footprint of approximately 100MB, which minimizes the need for distinct x64 CPUs. By transforming conventional storage servers into high-performance configurations powered by MinIO and specialized networking interface cards (NICs), AIStor optimizes resource usage within AI infrastructures.
Streamlining Inference Management
With the integration of NVIDIA NIM microservices, AIStor accelerates the deployment of inference infrastructures, allowing businesses to focus on strategic AI efforts rather than manual data management. The optimized model inference results from using NVIDIA hardware further boost the overall performance of the system.
MinIO AIStor is currently available for beta testing among select customers, with support for NVIDIA GDS and the integrated BlueField-3 networking platform expected to align with NVIDIA’s general availability timeline.