MinIO, a provider of high-performance object storage for AI, announced several upcoming enhancements to its AIStor product at NVIDIA GTC. These updates are designed to deepen MinIO’s support for the NVIDIA AI ecosystem and improve the efficiency and utilization of AI infrastructure. The new integrations are focused on streamlining data management and freeing up resources for AI-related tasks.
New MinIO AIStor features and integrations
MinIO announced three key advancements for AIStor:
- Support for NVIDIA GPUDirect Storage (GDS) for object storage: This integration establishes a direct data path between MinIO AIStor and NVIDIA GPU memory, bypassing the traditional CPU route. It will increase CPU efficiency. It also utilizes Ethernet networking fabrics, offering flexibility and cost-effectiveness for scaling AI infrastructure.
- Native integration with NVIDIA BlueField-3 networking platform: AIStor becomes the first and only object storage software to run natively on NVIDIA’s BlueField-3 Data Processing Unit (DPU). This is enabled by AIStor’s small footprint (~100MB). This eliminates the need for separate x64 CPUs, transforming storage servers into MinIO and NIC-powered JBOFs (Just a Bunch of Flash). It leverages Arm’s Scalable Vector Extension (SVE) instruction set for performance. It’s Spectrum X ready.
- Incorporation of NVIDIA NIM microservices into AIStor promptObject inference: This integration aims to simplify the deployment and management of inference infrastructure. AIStor’s new S3 API “promptObject” allows users to interact with unstructured objects using natural language, similar to interacting with a large language model (LLM). NVIDIA NIM model optimizations for NVIDIA hardware are intended to accelerate promptObject inference results.
MinIO’s co-founder and co-CEO, AB Periasamy, emphasized the company’s close alignment with NVIDIA, allowing them to innovate in AI storage at multi-exabyte scale. He highlighted the delivery of high-performance object storage on commodity hardware, enabling enterprises to maximize GPU utilization and reduce costs.
The company also claims that it created, a streamlined, ultra-fast pipeline that turns data lakes into high-speed AI/ML training environments.
Alex Timofeyev, Director, High Performance Compute Engineering and Operations at Recursion, mentioned preliminary testing suggests that MinIO AIStor will increase CPU efficiency in their AI compute infrastructure.
These new features are currently available to beta customers under private preview. General availability of MinIO AIStor support for NVIDIA GDS and native integration with the NVIDIA BlueField-3 networking platform will align with NVIDIA’s GA release schedule.
Featured image credit: MinIO
