F E A T U R E

– geophysics use cases face similar storage requirements, with a modern 3D seismic marine survey generating up to 10 petabytes of data for analysis, while in the financial services sector, fraud detection systems must process billions of transactions in real-time to prevent losses that can run into millions of dollars per incident. contemporary use cases, where databases and enterprise applications have created, in general terms, predictable sequential workloads.

As a result, and in contrast to the demands placed on storage by AI, organisations have benefitted from a greater degree of predictable scale and timing around their storage provision strategies. In the enterprise IT context, for example, a payroll system might process transactions in overnight batches, with relatively uniform demands on storage. Reliability also matters, but for pre-AI storage use cases, access patterns are relatively uniform and don’ t involve thousands of concurrent GPU processes hammering storage at once.

The arrival of advanced AI systems on the scale currently being seen is, however, game-changing. Training AI models is dependent on systems being able to read from massive, unstructured datasets( such as text, images, video and sensor logs, among many others) that are distributed and accessed in random, parallel bursts.

Instead of a few apps making steady or relatively predictable requests, a business might be running tens of thousands of GPU threads, all of which need storage that can deliver extremely high throughput, sustain low latency under pressure and handle concurrent access without becoming a bottleneck. When these requirements are not met, the impact on performance and cost can be immediate and severe.

For example, training a large language model can involve petabytes of text and image data being streamed to thousands of GPUs in parallel. If storage cannot feed that data at the required speed, the

GPUs sit idle – burning through compute budgets that can run into millions of dollars for a single training run.

In contrast, inference workloads create a different kind of stress. In financial services, for instance, AI-driven fraud detection must analyse billions of transactions in real-time, often within milliseconds. That requires storage systems capable of providing ultra-low latency access to massive historical datasets so that the model can compare live activity against years of prior patterns without delay.

The underlying point is that the legacy approach to storage is fundamentally unsuited to these performance extremes.

In with the new

For organisations reliant on HPC architectures, this is a well-trodden path. In the life sciences sector, for example, research organisations need uninterrupted access to genomic datasets that are measured in the petabytes. A great example is the UK Biobank, which claims to be the world’ s most comprehensive dataset of biological, health and lifestyle information. It currently holds about 30 petabytes of biological and medical data on half a million people and is among many similar projects around the world gathering information at a tremendous rate.

In government, federal agencies face their own challenges in supporting programmes that cannot afford downtime or data loss. Mission-critical applications, such as intelligence analysis and defence simulations, demand 99.999 % uptime, and even brief interruptions in availability can potentially compromise security or operational readiness. The list goes

Suffice to say that in each case, storage capacity and resilience have become

36 www. intelligentdatacentres. com