The Data Centre Revolution Demands Sustainable AI

As AI demands greater levels of power, the data centre industry is set to be impacted by the increasing need to be energy-efficient.

With data centres currently contributing 2% to the world’s GHG emissions, facilities are producing electronic waste too - which accounts for 70% of toxic waste and 2% of solid waste, according to research by 8 Billion Trees.

As a result, new innovations are emerging within the sector as a way to combat rising power uses. To discuss some of these, Data Centre Magazine speaks with Tim Rosenfield, co-CEO of Sustainable Metal Cloud & Firmus Technologies. He examines the impact of AI on the sector and how businesses can harness the technology to boost their sustainability strategies.

Sustainable Metal Cloud is the GPU customer offering of Firmus Technologies.

How are you currently seeing the demand for AI technology impacting the data centre sector?

There has been a tectonic shift in the technical nature of computing from CPUs to high-powered GPUs, forming large platforms or clusters known as AI factories. These AI factories support the explosive growth in generative AI (Gen AI). To meet the needs of the existing and future generation of AI chips, the energy required will double by 2030. 

According to the IEA, data centres globally will double electricity consumption by 2026 and potentially quadruple by 2030. Much of this data-centre growth is driven by new applications of Gen AI.

The central impact of all this growth has been to force a discussion on the right way to build these AI factories, from an energy efficiency, capability (density) and TCO perspective. Over 95% of the world’s existing data centre fleet is not ready for liquid cooling - a must have technology to build these factories. The question is, what is the right platform to upgrade these facilities, and what will lead to the best cost outcome to build, operate and open the door to a future proof platform for years to come?

What are the biggest challenges that businesses face when it comes to implementing AI for sustainability initiatives?

A common misconception is that cost is the biggest challenge in implementing AI for sustainability initiatives. In reality, sustainable solutions like our immersion cooled AI computing platform save on energy and space, and therefore cost. The technology behind our approach reduces energy consumption by up to 50% compared to traditional air cooling, making it a cost-effective solution for high-performance AI workloads.

Our challenges are not technological but lie in addressing technical scepticism, economic cynicism, and resistance to changing the status quo. We overcome technical scepticism through rigorous independent benchmarking, demonstrating the efficiency and performance of our technology.

What are some of the ways that AI can help data centres achieve their sustainability goals?

AI-driven predictive analytics can optimise energy usage by adjusting cooling systems and managing workloads more efficiently. Machine learning algorithms can monitor and predict equipment failures, reducing downtime and maintenance energy costs. AI can aid in designing more efficient data centre architectures and managing energy consumption across the facility. 

All of this happens from within existing data centres - even if they were never built for liquid cooling to begin through our core design tenants of building for hyperscale, and solving for interoperability. 

We build for hyperscale by going large. Ours are 1MW self-contained modules, comprising 16 tanks of 52RU, connected in series to a central plant room - a unique approach in the industry. This design maximises efficiency by employing mechanical economies of scale, whilst vastly improving modularity and scale - each 1MW ‘building block’ (with current OEM chassis designs can hold up to 768 H100, H200 or B200 GPUs, but can go denser with newer designs) scales linearly, enabling the creation of a truly straightforward process to build AI factories.

For interoperability, this is where design innovation meets efficiency. Our platform’s immersion cooling loop is self-contained, yet the way we handle heat rejection outside of the facility is unique. Our platform requires no chillers, refrigerants or other high-energy cooling systems from the building, unlike most other liquid systems, including direct chip cooling. 

These innovations are just some of the design principles that have contributed to our platform emerging as the most energy and cost efficient method to build hyperscale AI factories. Whilst other liquid technologies like direct-to-chip can solve for capability (rack density), they don't necessarily solve for energy efficiency, operating cost or capital cost advantages against a tightly integrated approach such as ours.

Do you think that government regulations will help to drive data centre sustainability?

Government regulations can significantly drive data centre sustainability. 

Countries like Singapore, Ireland and the Netherlands have been the first to recognise these issues and have limited data centre development and grid access. Singapore has focused on sustainability criteria for land and power allocation, to address the growing energy draw from their grids to support data centre operations. It’s why we decided to base our business in Singapore. 

We need a balance between government input, control, and regulation, which mandates that capital deployment not only generates returns for investors but also benefits the local community and the environment.

How can data centre businesses work to deliver sustainable, cost-effective AI computing?

The uncomfortable reality is that the vast majority of data centres built today are not fit for purpose when it comes to delivering for large-scale AI workloads - let alone doing this cost effectively or sustainably.

If liquid cooling is the essential ingredient to achieving this objective, then the question becomes how to deliver this. What is an optimised platform, not just for now, but for future generations of AI accelerators, and what is the balance to strike between solving for technical capability against total cost of ownership.

Our thesis on this is clear. The world's data centre operators are now looking to retrofit with liquid technology. 

Hybrid cooling approaches like direct-to-chip are a hybrid vehicle. Direct-to-chip does solve for capability (rack density), but owing to complexity and design it is not the most cost effective to build or run. Direct-to-chip environments still require fans in the server, and in most cases chilled water. This means that over air cooling there is a proportionally small gain in overall PUE and energy efficiency. Operating costs through power consumption are not improved as much as they could be, and owing to complexity this is also a quite costly solution to build. 

Data centres today underpin much of the global economy, supporting everything from cloud computing and AI processing to the digital infrastructure necessary for modern business operations and everyday activities. As their role only grows in importance, it’s important to continuously push for better, more sustainable practices and ultimately reduce the environmental footprint of DCs throughout this fourth industrial revolution and beyond. 

******

Make sure you check out the latest edition of Data Centre Magazine and also sign up to our global conference series - Tech & AI LIVE 2024

******

Data Centre Magazine is a BizClik brand