By President Ntuli, Managing Director of HPE South Africa

Some have said generative AI is the defining technology of our time. It has the potential to advance the way people live and work on a scale unlike any technology we’ve seen to date. For South African businesses, the potential of AI is highly appealing, promising to usher in a new era of innovation and efficiency. In fact, the AI market in South Africa is projected to grow at a significant annual rate of almost 30% over the next six years, reaching a market volume of $4 billion by 2030. This surge in growth signals a revolutionary shift towards AI technologies across various sectors in the country.

Storage is a critical element of any AI architecture — but one that doesn’t get enough attention. What storage environment is suitable for enterprise-wide AI adoption? Is there a single storage architecture that can work for every AI initiative? How can you simplify data storage for AI?

The answer to all these questions is that it depends on your organisation and its priorities. But there is one truism that applies across the board: Any AI storage solution must be able to achieve scale in terms of performance, simplicity, and efficiency — regardless of the specifics.

For example, organisations focused on having the fastest file performance might opt for a distributed, parallel file system paired with a hybrid cloud-capable file and object storage solution that allows them to economically meet capacity needs.

Other organisations might want to extend their high-performance computer-based research environments for AI workloads, addressing high-demand input/output requirements with a smaller footprint and fewer drives. Still others may be looking to use a data lake, using AI to manage the heavy lifting involved.

The final, remaining organisations may simply use general-purpose enterprise file storage to store, manage, and curate active unstructured data on-premises and in the cloud, offering a consistent user experience wherever data resides, from edge to data centre to cloud.

As is often the case, one size doesn’t fit all, and no single solution will fit every AI requirement, maturity stage, and workload optimally across the board.

How to think about AI storage architecture

The four main considerations around AI storage are GPU utilisation, scalability, simplicity, and efficiency. Let’s look at each one in turn.

1. GPU Utilisation

Underutilisation of GPU resources will negatively impact cost and productivity. When an organisation makes expensive investments in a GPU farm, it needs to realise a return on that investment by using it as fully as possible. To do so, eliminating storage bottlenecks is essential. If the storage platform cannot keep pace with the computational power of GPUs, money is being wasted. Code and system setup must both be optimised to enhance GPU utilisation. Training AI models requires days, weeks, or even months, and wasted GPU cycles prolong the training even further. Underutilisation also means lower productivity among data scientists and data engineers and delayed time to market for new products.

2. Scalability

Storage must support all stages of AI, including data aggregation, data preparation, model training, model tuning, and inferencing. Your storage architecture needs to scale with that journey. It also must scale as models become more sophisticated and data volume increases. Traditional network attached storage (NAS) devices will quickly hit a performance ceiling that prevents them from keeping up, meaning they won’t work for most AI environments.

Scalability is key because in AI, the amount of data only gets larger and larger. AI isn’t a onetime training process. As new data is generated, existing models may need to be retrained with the new data to improve their performance. That means storing not just the current training data but a large amount of historical data too — a never-ending growth cycle that requires boundless scalability with room for more and more capacity as AI needs evolve.

3. Simplicity

All the above storage needs can get complicated quickly, overwhelming IT and data science staff. Simplicity is therefore key: Infrastructure and data management must be kept as simple as possible, no matter how many petabytes of data are needed. The simpler your storage system is, the more productive IT and end users will be.

4. Efficiency

Storage solutions must also leverage effective data reduction, minimise physical data centre space, and provide efficient power and cooling to reduce carbon footprint and keep sustainability high.

What type of storage is best?

Object storage is ideal as a scalable and cost-effective solution for storing large volumes of unstructured data, which is common in AI applications. Block storage is great for low-latency, high-performance access to data, making it well suited for AI workloads that demand rapid data processing. And file storage is a solid fit for AI and machine learning workloads that need low latency as they process large amounts of file data.

Which one you choose depends on your priorities, and unfortunately, there’s no one-size-fits-all approach to AI data storage. The best strategy will consider each organisation’s unique needs, data volumes, and budget — and it will likely change over time.

As AI technology keeps advancing, so too must our storage solutions. Keeping up with the latest developments means you'll always be able to get the most out of your AI applications. It's essential for South African businesses to stay ahead in the ever-evolving world of AI, and ensure we realise the full potential of the AI opportunity.