Mauro Leuce at Colt DCS explains how a massive increase in computing needs demands a rethink.
Artificial intelligence is having an unmistakably huge impact on industries from transport and agriculture to energy and finance. Over the past year, that has gained an extra dimension as generative AI has made a splash across the world.
There are big predictions on how much this market might explode over the coming years. For example, a report by Bloomberg Intelligence last year forecast that the generative AI market will grow from just $40 billion to $1.3 trillion in the decade to 2032.
Such enormous anticipated rises give the data centre industry pause for thought on how to align strategies with the massive increase in computing needs. “For hyperscale data centre operators, we believe this will result in new opportunities,” says Mauro Leuce, global head of design and engineering at Colt Data Centre Services (DCS). “But it also demands a rethink of how we meet customer requirements with the most effective data centre design and construction.”
Indeed, the early indications from what Colt DCS has seen are that it will have a major impact, says Leuce. “There’s been a surge in customer requests in terms of us being ready to accept these HPC [high-performance computing] workloads,” he says.
Powering up
As demand grows, one thing that AI requires is for data centres to provide a lot more power per square metre in data halls. While Colt DCS has seen a peak power demand of around 20kW for standard workloads per rack, this has risen to 130kW-plus in the last couple of quarters for AI workloads. Although that peak requirement is still relatively niche, the company believes demand for that amount may become much more widespread in the next two years.
“We’re very capable of helping our customers to define bespoke solutions to meet their demands,” says Leuce.
He explains that Colt DCS has been preparing for some time for a ramp-up in demand for high-power workloads, making it well-prepared to cater for the AI surge. As part of this, it carried out a review of three major areas in 2020 – water use, heat reuse and HPC.
That has resulted in changes, including measures such as a hybrid cooling solution based on air and liquid cooling to deal with increased heat generated by greater rack density and to improve sustainability. “Colt DCS’s Global Engineering Team had to fully review the reference design for our new-build data centres, with a focus on increasing its flexibility to accommodate new and hybrid customer requirements, while targeting environmental and operational concerns like water use and heat reuse,” says Leuce.
“All our new data centres in Paris, Frankfurt, London, Tokyo and Mumbai have been or will be delivered following the new reference design.”
Playing it cool
When it comes to cooling, increased power demand has led data centre operators to look more towards liquid-based systems, says Leuce, because servers can no longer be cooled down simply by transferring heat to air. These improve efficiency by transferring heat directly from the source to liquid, which has a higher heat capacity than air. As a result, explains Leuce, Colt DCS started to steer away from air systems using indirect evaporative cooling a few years ago to focus on chilled-water systems, which give flexibility to accommodate hybrid systems using a combination of air and liquid cooling.
Chilled-water systems have environmental benefits too, delivered through their ease of connection to district-heating systems. Using these prevents energy being lost in the environment by allowing waste heat from data centres to be reused in commercial and residential buildings. This then contributes to energy strategies set by governments to reduce gas demand and emissions.
Another consideration for data centre design and deployments is that AI creates two different types of workload, one for training AI models and another for ‘inferencing’ – or putting trained models into practice. Leuce says this leads to a need to consider different service-level agreements (SLAs) for infrastructure availability in data centres. These could comprise the inclusion of backup generators, uninterruptible power supplies and continuous cooling to retain business continuity for inferencing workloads, but less-demanding SLAs on this front for training workloads.
Furthermore, it should be considered how best to set out data halls to minimise customer costs for network cabling by providing the shortest routes between racks used for AI to address latency challenges.
Retro challenge
An additional significant challenge with AI is in retrofitting facilities to deal with the transforming demands. Available IT power may be limited by various factors, including the available power of existing utility connections, and local rules on air and noise emissions for backup generators.
Through the wider-scale adoption of liquid cooling, however, Colt DCS has the potential to offer extra IT load to customers through an improvement in power-usage effectiveness and increased white space from higher power densities.
But one of the biggest obstacles comes in finding space for retrofitting chilled-water systems, which may be difficult in already congested facilities located in city centres, even though there are options such as installation on building roofs or side gantries. Colt DCS has, nonetheless, succeeded in accommodating customer requests for AI workloads even in heritage data centres, says Leuce.
Bigger capacities
At the same time, the company has been building in much higher power capacities at its new locations than the traditional 20 to 30MW levels of earlier ones, offering customers fresh options for coping with the big demands of AI and for planning ahead for the future. For instance, the Mumbai facility that Colt DCS opened last year as its first data centre in India will be capable of supporting 120MW-plus of IT power when the entire campus is developed, with two more facilities in India and Tokyo planned to deliver 70MW of IT capacity. Meanwhile, its new data centre campus in Hayes, West London, will support up to 150MW.
Such facilities could, in particular, offer big benefits for inferencing workloads, which need to be as close as possible to users, while hyperscale operators may choose to keep less-demanding training workloads in their own facilities. For operators that find the take-up of AI outpaces their ability to keep up with growth, the facilities of Colt DCS may also provide an attractive option.
Although it remains to be seen exactly what the ultimate impact of AI will be, Leuce says Colt DCS has managed to prepare itself for the potential surge by planning ahead and having visibility into the market through its close relationships. “This comes from being a trusted partner, offering quality products delivered on time and to budget with competitive rates,” says Leuce. “That has created stronger relationships with customers, which now see us not just as a provider, but as a partner of their own growth strategies.”
This means Colt DCS has already been able to design data centres to incorporate the latest needs for liquid and hybrid cooling, as well as to be ahead of the game on measures like district heating and zero water usage, he adds.
Sustainable path
But that’s just part of how the company is improving sustainability as it deals with the AI influx. In addition, things like new ‘active-active’ HPC architectures will enable the omission of backup generators, instead relying on battery storage from uninterruptible power supplies to protect critical loads during outages.
“This will be reflected in less carbon emissions due to less fuel being consumed during data centre operations, but, more importantly, less embodied carbon, with a reduction when it comes to building structures, and the fabrication and shipping of generators and fuel tanks,” says Leuce.
On top of that, he points out that Colt DCS will use AI tools themselves to optimise efficiency in data centres, further reducing environmental impact. “The changes required to adapt our old and new facilities will only contribute to the success of our environmental, social and governance strategy,” says Leuce.
When coupled with the company’s market knowledge, that creates a strong data centre proposition for customers, he adds. “For me, some of the key advantages of Colt DCS are that it has a presence in the right locations, with the right power levels,” says Leuce. “And, of course, we continue working closely with our customers to identify new locations, and when to procure power and connectivity.”
