In this podcast, we sit down with Tim Sherbak, who manages enterprise products and solutions at Quantum. We dive into how artificial intelligence (AI) is reshaping data storage, especially when it comes to storing large volumes of data over extended periods.
We discuss the specific technical needs that AI demands from storage systems. For starters, AI relies on all-flash storage in scalable architectures and needs robust data throughput, both across multiple streams and individual data flows. Tim emphasizes the concept of “forever growth” and “forever retention,” which challenges organizations to rethink how they manage and optimize storage to handle endless data expansion. He highlights the FAIR principles—findability, accessibility, interoperability, and reusability—that emerged from the scientific community as a way to manage data more effectively and openly.
We also explore how storage providers can use AI to tackle the massive quantities of data spread across diverse environments.
Let’s take a closer look at what AI means for data storage. AI processing demands a lot from data storage systems. Neural networks consume vast amounts of data and require high computational power. The main challenge? Feeding these hungry systems efficiently. The goal is to ensure that powerful GPU clusters operate at full capacity since they need vast data for processing. This requires high throughput and low latency.
To meet these demands, organizations are turning to NVMe and all-flash solutions. These technologies are designed to scale seamlessly, enabling large clusters to function optimally. It’s crucial for all compute clusters to have access to data in a flat namespace, ensuring visibility across the board.
Right now, a lot of attention is given to RDMA, or remote direct memory access, which lets servers and storage nodes directly tap into storage resources. This setup enhances storage access throughout the cluster. Moreover, it’s not just about aggregate throughput; single-stream performance is equally critical. New architectures now allow for parallel data paths, optimizing data delivery to the GPUs.
Now, how can organizations manage storage more effectively in light of AI’s demands? Two major challenges stand out. First, there’s the endless growth of data. Second, we need to think about long-term retention. The scale of data generated far exceeds what’s manageable from individual GPU runs.
This data must be preserved affordably over the long haul. Some solutions combine flash, disk, and tape storage to optimize both cost and performance. This hybrid approach lets organizations tailor their storage strategies to fit budgetary and performance needs.
Another recommendation for tackling the challenge of endless data is to utilize FAIR data management principles. Emerging about six to eight years ago, this concept originates from research organizations looking to curate data. It also holds real potential for improving AI dataset management. FAIR stands for findable, accessible, interoperable, and reusable. By aligning data management strategies with these principles, organizations can enhance their data infrastructure and maintain better oversight.
Finally, let’s discuss how AI can aid in data storage management. The potential here is fascinating. As storage vendors gather insights from customer data, they can improve infrastructure support on a global scale by analyzing usage patterns. However, the most intriguing application is the idea of self-aware storage, or self-aware data management.
This involves cataloging rich metadata about stored data, with AI handling the heavy lifting of cataloging and pattern recognition. As datasets grow, AI will automatically classify and document this information, making it easier for organizations to utilize their data. For instance, consider sports: AI could efficiently catalog a player’s entire career by analyzing footage, articles, and other resources. When it’s time to highlight a player after retirement, instead of a frantic search for data, AI would streamline access to all relevant content.