NVIDIA’s RDMA for S3 Ushers in new Era of AI Storage Performance
A new approach to object storage, leveraging Remote Direct Memory Access (RDMA), promises to dramatically accelerate AI workloads and reduce costs as data demands skyrocket.
Today’s artificial intelligence applications are insatiable consumers of data, demanding storage solutions that are both scalable and affordable. Projections indicate enterprises will generate nearly 400 zettabytes of data annually by 2028, with a staggering 90% of this new data being unstructured – encompassing audio, video, PDFs, and images. This exponential growth,coupled with the need for seamless data portability between on-premises systems and the cloud,is driving the AI industry to explore innovative storage options.
RDMA for S3: A Performance Leap
Enter RDMA for S3-compatible storage, a solution utilizing remote direct memory access (RDMA) to accelerate the S3-request programming interface (API)-based storage protocol. Optimized specifically for AI data and workloads, this technology represents a important departure from customary storage approaches. While object storage has long been favored for cost-effective storage of data like archives, backups, and activity logs, its performance limitations have historically hindered its use in demanding AI training scenarios.Now,that’s changing.
This new solution, built around NVIDIA networking, delivers faster and more efficient object storage through RDMA-enabled data transfers. The benefits are substantial: higher throughput per terabyte of storage, increased throughput per watt, reduced cost per terabyte, and significantly lower latencies compared to TCP, the conventional network transport protocol for object storage.
Key Advantages for AI Enterprises
The advantages of RDMA for S3-compatible storage are numerous. It enables faster data access for AI training and inference, leading to reduced model development times and improved application performance. The technology also enhances storage utilization, allowing organizations to maximize the value of their existing infrastructure. furthermore, the open architecture fosters innovation and collaboration, ensuring that the solution remains adaptable to evolving AI needs.
Importantly, the architecture is designed to be open, allowing other vendors and customers to contribute to the client libraries and develop their own software to support and utilize the RDMA for S3-compatible storage APIs.
Industry Collaboration and Standardization
NVIDIA is actively collaborating with partners to standardize RDMA for S3-compatible storage, paving the way for widespread adoption. Several leading object storage providers are already embracing the technology, including Cloudian, Dell Technologies, and HPE.
“Object storage is the future of scalable data management for AI,” stated a chief marketing officer at Cloudian. “Cloudian is leading efforts with NVIDIA to standardize RDMA for S3-compatible storage, which enables faster, more efficient object storage that helps scale AI solutions and reduce storage costs.”
Dell Technologies echoed this sentiment, with a company spokesperson noting that, “AI workloads demand storage performance at scale with thousands of GPUs reading or writing data concurrently, and enterprise customers…desire AI workload portability for objects.” Dell has integrated RDMA for S3-compatible storage acceleration into its Dell ObjectScale platform, delivering unmatched scalability, performance, and reduced latency.
Similarly,HPE emphasized the transformative potential of the technology. “NVIDIA’s innovations in RDMA for S3-compatible storage APIs and libraries are redefining how data moves at massive scale,” said a senior vice president and general manager of storage at HPE. “Working closely with NVIDIA, HPE has built a solution that accelerates throughput, reduces latency and lowers total cost of ownership.” HPE has integrated RDMA for S3-compatible storage into its HPE Alletra Storage MP X10000 platform.
Availability and Future Outlook
NVIDIA’s RDMA for S3-compatible storage libraries are currently available to select partners and are slated for general availability through the NVIDIA CUDA toolkit in January. Furthermore, NVIDIA is introducing a new NVIDIA Object Storage Certification as part of its broader NVIDIA-Certified Storage programme. As AI continues to evolve and data volumes surge, RDMA for S3-compatible storage is poised to become a cornerstone of modern, high-performance AI infrastructure.
