PEAK:AIO talks about optimising storage for AI

In many ways, the AI-driven revolution is supported by software. In order for hardware like GPUs to run AI workloads and do so efficiently, the software needs to be sufficiently optimised.

PEAK:AIO understands this, with a heritage centered on AI that now focuses on software-defined storage and helping customers “to get the most out of their GPUs”, as Lanigan termed it, looking at what its customers do and don’t need.

“There are lots of companies out there who are building  massive, big storage solutions, but for so much of the AI market, they’re … overkill in terms of price, in terms of size, in terms of the energy consumption,” explained Lanigan. “So we decided … let’s build something from the ground up that’s just software only, and work with off-the-shelf hardware, because people like to have options.” 

Optimising the software means the hardware can be optimised too, in the form of having fewer GPUs and smaller clusters, which results in improved energy efficiency and reduced costs - both big boons for companies who are aware that investment in AI can be significant.

“One of the oldest tricks in the storage book has been, if you need more performance, data throughput to deliver to compute, you just add more hardware,” said Lanigan. “You add more physical hardware that adds a huge amount of cost and it adds complexity. We wanted to avoid that and wanted to serve those customers.” 

There can be an assumption that in order to integrate AI, companies need lots and lots of GPUs to do so - but this doesn’t have to be the case.

“You … often see that where first a customer will say, ‘Okay, we’re looking for a SuperPOD (a massive cluster of NVIDIA GPU servers)’ but it turns out that they might just start with a really small fraction of that, and then eventually they might go to that, but they can do so much more than they realise, if they just optimise the infrastructure that they have.”

A traditional mindset is to directly correlate what performance is needed with the cost.

“But like every industry, it needs a disruptor, and it needs someone to come along and say, ‘We need to stop thinking about things in the old way’,” continued Lanigan. “That’s what our customers have told us … and realistically, it’s our responsibility as the vendors to think … how can we do things in a more efficient way so that we’re not passing the cost of inefficiency onto the customer.” 

Optimising software-defined storage 

What are PEAK:AIO doing differently?

By challenging the traditional means of improving performance in storage - i.e., adding more hardware - PEAK:AIO focused on creating a single-server solution that transformed into a “building block”.

“We’re not dealing with the limitations of legacy design,” he said. “That’s been the problem of the storage industry over the last number of years … where storage has developed in a certain way, and vendors have iterated upon these previous iterations, and then you just get these kinds of clunky solutions … where, in AI, these features are supported that customers don’t even need, but they’re still paying for.” 

One solution showcasing the company’s capabilities at the AI Summit London was the Boston Igloo AI+, a 1U all-flash appliance with 500K IOPS. It was introduced in 2023 and designed to meet demand for ultra-fast AI storage solutions.

This demonstrated not just PEAK:AIO’s agnosticism to hardware solutions, but their focus on companies that can provide customers single-node solutions.

“That’s great for companies who are cost-conscious or energy-conscious, and they want to have … something super simple to manage what they call a base part of GPU servers (small to mid-size cluster of GPU servers),” said Lanigan. “Or it can serve Edge solutions really well … that’s interesting for the IoT world, where people are deploying solutions at the Edge.”

It also demonstrated rack-scale solutions with its partner Western Digital, where it acted as the software management layer controlling storage and data, and could be scaled from 100 terabytes to 10s of petabytes.

An example of where PEAK:AIO’s software-defined storage has been deployed includes in hedgehog conservation - placing camera traps around London to identify what’s a hedgehog and what might be another animal, for example.

“All this information gets sent back to a central data centre, which is … like a little container at a zoo. And there was a really small amount of space for this to go into. Most of the space was taken up by these massive … GPU systems,” explained Lanigan. “But the customer said, ‘well, we still need about three petabytes of data’.

“Any other vendor would say, well, you need a serious amount of hardware for that. We did it in a small solution … with an enormous amount of performance.”

This shows how the “help of AI” can be applied to real-world problems, and how focusing on optimising software-defined storage can aid issues like hedgehog conversation - as well as save companies space and cost.