Once limited to academic institutions and scientific research projects, AI is now being explored for enterprise purposes such as streamlining supply chains or enhancing customer experiences. But in most cases, AI still needs to be widely used across the entire organisation, limiting its benefits.
Legacy IT infrastructure is among the top barriers preventing the pervasive use of AI. Also, research by the International Data Corp found that inadequate or lack of purpose-built infrastructure capabilities often cause AI projects to fail.
Agreeing with that finding, Pure Storage’s CEO Charlie Giancarlo asserts that the conventional hard disk drive (HDD) — which consists of a spinning disk to read and write/store data magnetically — will not be able to cope with the demands of an AI-driven organisation. “HDDs used to be in everything, but they’re no longer in your laptop, handphone and more. Today, enterprise mass storage, data centres and hyper scalers are the last remaining bastions of hard disks,” he said in his opening keynote at last month’s Pure//Accelerate 2023 conference in Las Vegas. He also boldly claims that no new HDDs will be sold by 2028 as they will be entirely replaced by flash storage, whose capabilities have been rapidly developing over the years to address the needs of AI.
AI’s requirements
At the same conference, Pure Storage’s chief technology officer Rob Lee shares the IT infrastructure (particularly from storage and computing) requirements for AI.
The first is performance. He explains that AI is a game of “who can collect the most data, feed it into graphics processing units (GPUs) the fastest [for analysis], and repeat the process the most”. Therefore, the infrastructure must enable AI to perform all those tasks efficiently and across different data types such as text, images and video.
See also: 80% of AI projects are projected to fail. Here's how it doesn't have to be this way
He also highlights that the data collection part of AI is a commonly overlooked area for ensuring the technology’s performance. “Back in the day, it was said that about 85% of the work and time spent in data science was on preparing the data for the analyst. I’m not sure what that looks like today with AI, but it is considerable. So, when you assess AI’s performance, you should also consider the responsiveness of all the systems that collaborate to collect data and bring it into your back-office systems. The systems that manage data security, index and label data, and more, matter too,” he says.
Secondly, AI calls for a flexible infrastructure. “Since the tools, techniques, datasets and algorithms we use are changing daily, you need to invest in an IT infrastructure that will enable you to adapt and evolve. If you lock yourself into a technology stack that’s too inflexible, you’ll get left behind,” he says.
Reliability is another requirement as AI is increasingly being used in mission-critical environments. Lee says: “As organisations increase their use of AI, they will use more power-hungry GPUs. But the power supply is declining, and the availability [of alternative power sources] is not growing quite as quickly. So, you’ve got to invest in a more efficient IT infrastructure.” He adds that the infrastructure should also help address the privacy and security concerns around AI by providing security controls and data governance capabilities.
See also: Responsible AI starts with transparency
Lowering AI’s carbon footprint
While AI can help businesses in many ways, it may come at a cost to the environment as it leaves a large carbon footprint. According to MIT Technology Review, training just one AI model can emit more than 626,000 pounds (283,900 kg) of carbon dioxide equivalent, nearly five times the lifetime emissions of an average American car.
One way organisations can reduce AI’s carbon footprint is to deploy high-density storage solutions that consume less power, maximise capacity utilisation and offer an extendable life. “[This is why Pure Storage’s flash storage solutions are designed to] deliver 10 times more power and space efficiency than HDDs. Our solutions are also two to five times more power- and space-efficient while being 10 times more reliable than competitors. It also requires five to 10 times less manual labour than operating legacy storage. Adding all those things together will result in at least 50% less total cost of ownership compared to competitive offerings of both flash and HDDs,” claims Giancarlo.
Additionally, Pure Storage’s Evergreen model and services allow organisations to buy its flash storage solutions once and use them for a long time. The subscription service, says Giancarlo, allows storage to be scaled and upgraded without causing downtime or negatively impacting performance. This also prolongs the life of Pure Storage’s flash storage solutions to 10 years or more. Case in point: More than 97% of Pure Storage’s FlashArrays over five years old are still in everyday service. As such, the Evergreen subscription service helps improve flash storage’s agility and longevity, reducing e-waste.
Using flash storage as AI’s foundation
Meta is one organisation that leverages flash storage for its supercomputer. Called AI Research SuperCluster (RSC), the supercomputer is helping Meta’s AI researchers build new and better AI models that can learn from trillions of examples; work across hundreds of different languages; seamlessly analyse text, images and video together; develop new augmented reality tools; and much more.
With Pure Storage’s FlashArray and FlashBlade, RSC can rapidly analyse structured and unstructured data and direct more power to GPUs as the flash storage solutions consume low power. RSC also reduces overall operational costs as Pure Storage’s solutions provide reliable performance and require less maintenance than disk storage.
To stay ahead of the latest tech trends, click here for DigitalEdge Section
Closer to home, South Korea’s MediaZen uses Pure Storage’s FlashBlade to accelerate time to market for new AI services and enhance its R&D capabilities. The company provides voice recognition solutions to automotive, education, public service, retail and telecommunications industries.
With FlashBlade, MediaZen reduced voice recognition modelling tasks that took up to 12 months to two weeks. New speech recognition models were created in four weeks with high-speed shared storage supporting a multi-GPU distributed processing environment.
MediaZen also created a new infrastructure environment with FlashBlade to advance its AI R&D and develop and expand services specific to market needs. This has enabled it to offer diverse markets with AI-powered voice and language services, ultimately expanding its global footprint.
“With FlashBlade, MediaZen has the infrastructure to advance its AI services to meet current and future market demands. We are thrilled with the superb performance of FlashBlade, and its simple operation and maintenance that requires no additional headcount to manage,” says Yoon JongSung, deputy director of the NAMZ AI R&D Group at MediaZen.
“The AI era has arrived, and the need for modern all-flash storage systems to support large-scale AI workloads and advanced data analytics is increasing daily. To support market leaders, including MediaZen, to innovate and grow their businesses, Pure Storage will deliver higher performance, capacity density and reliability with our differentiated flash technology,” says Jaesung Yoo, managing director of Pure Storage Korea.
