“Responsible AI by Design”: Practical sustainability considerations in adopting Gen AI
Written by Jens Eriksvik & Michael Asplund
AI offers significant opportunities for innovation and efficiency. However, alongside these advancements it is important to ensure AI is developed and deployed responsibly. We have all heard about “by design”-approaches, and now is the time for "Responsible AI by design". This approach mitigates risks, reduces long-term AI model maintenance costs, and builds trust with stakeholders. It is also key to reducing the environmental impact of AI.
"Responsible AI by Design" (RAID) is a comprehensive framework and methodology that systematically integrates principles of transparency, fairness, accountability, and sustainability throughout the entire lifecycle of AI development and deployment. Responsible AI ensures that AI systems are designed to mitigate risks, build trust, and maximize positive societal impact from inception. By embedding these principles into AI strategies, organizations prioritize responsible practices while navigating the complexities of AI technologies adeptly.
“At Algorithma, we believe that the key to successful AI deployment lies in strategic planning and sustainable practices. By embedding Responsible AI by Design principles into our AI strategies, we not only mitigate risks but also enhance the overall trust and value of our AI solutions for our clients."
- Jens Eriksvik, CEO at Algorithma
From a sustainability perspective it is clear that the advent of Gen AI has led to a surge in energy consumption. Morgan Stanley forecasts a 70 % YoY increase until 2027. Consequently, businesses are rushing to secure access to green/sustainable energy, e.g. as Microsoft’s recent data center investment in Sweden highlights.
In a recent article Forbes deep-dives into the problem of surging power consumption from Gen AI. Big Tech's investment in AI accelerators, like Nvidia's A100 and upcoming generations, is driving exponential growth in power consumption. Each new generation of GPUs increases performance but also escalates power demands significantly. With forecasts indicating a tripling of global data center power use and projections of AI-related electricity demand skyrocketing, sustainability becomes a critical concern. Businesses are increasingly adopting green electrical power and energy-efficient technologies to mitigate environmental impacts. The industry is actively addressing the power challenge through technological innovations and strategic investments in renewable energy, BUT it is estimated that “Nvidia’s 3.76 million GPU shipments [in 2023] could consume as much 14,384 GWh (14.38 TWh)”. The 14.4 TWh is equivalent to the annual power needs of more than 1.3 million households in the US - and this is only counting 2023 Nvidia deliveries (Forbes; no AMD, no BigTech, no other stuff).
“By working with renewable energy sources, monitoring and optimizing energy consumption, and building dense computing environments, we ensure that our AI operations are both cutting-edge and environmentally responsible. However, optimizing data center operations and a modern cloud platform for sustainability is not just about reducing costs; it's about working towards a sustainable future."
- Michael Asplund, Operations manager at Elastx
Thus, putting the energy consumption issue aside, leveraging AI involves several other key areas of responsibility. From a regulatory standpoint, it's crucial to comply with data protection laws, intellectual property rights, and industry-specific regulations that may govern AI use. Ethically, organizations must consider the potential societal impacts of their AI systems, including issues of bias, fairness, and transparency, to ensure AI is not used for harmful or manipulative purposes. Additionally, as AI becomes more prevalent, businesses should be prepared for evolving corporate processes and business model structures. RAID also means investing in robust security measures to protect AI systems and the data they process.
“Elastx's wholehearted commitment to our staff, customers and society involves delivering sound products with excellent support, paying taxes, and contributing to the open-source community. This commitment is firmly built into our approach."
-Michael Asplund, Operations manager at Elastx
With a RAID mindset, there are four key aspects that need to be considered for businesses embarking on a generative AI journey: Gen AI deployment strategy and policy, data center optimization (or procurement thereof), AI architecture and model efficiency, and AI operations and AI model management.
Gen AI deployment strategy and policy
Review and strategically choose where to deploy Generative AI solutions to maximize efficiency and impact, especially see where Gen AI actually can be used to offset environmental impacts (net positive approach)
Evaluate whether AI deployment is necessary in every scenario, prioritizing applications with the highest value and energy efficiency.
Consider edge computing to process data closer to its source, reducing the need for extensive data transmission and central processing.
Establish policies that prioritize energy efficiency in AI operations, and set energy consumption targets and regularly review performance against these benchmarks.
AI architecture and AI model efficiency
Based on deployment strategy, architect for RAID, scope the usage of Gen AI to distinct business use-cases with clear value - review what AI tech to deploy where
Employ techniques such as model pruning, quantization, and knowledge distillation to reduce model size and complexity, and optimize AI algorithms to run with lower computational requirements.
Use of energy-efficient programming languages and frameworks.
Data center optimization
Partner with suppliers that provides energy-efficient GPUs and AI accelerators, and, where applicable, make sure that specialized AI chips designed to minimize power consumption is used. Keep it fit for purpose.
At end of life, make sure that the supplying partner upgrade to energy-efficient data center infrastructure, implement advanced cooling systems and other technologies to reduce power usage and consolidate workloads to minimize the number of active servers.
Secure access to renewable energy sources (solar, wind, hydro) for powering AI operations, and partner with green energy providers to ensure a sustainable energy supply.
AI operations and AI model management
Utilize energy monitoring tools to track and analyze power consumption in real-time, and implement AI-powered analytics to identify and address inefficiencies in energy usage.
Use shared cloud services with dynamic scaling to adjust resources based on demand, thereby reducing energy wastage, and implement auto-scaling to shut down unused resources during low-demand periods.
Optimize data management practices to reduce the volume of data processed and stored, and use synthetic data or data augmentation techniques to minimize the need for extensive data collection.
Use workload scheduling to run energy-intensive tasks during periods of low energy costs or high renewable energy availability.
Conduct life cycle assessments to understand and mitigate the environmental impact of AI systems from development to deployment, and regularly assess and update AI systems to incorporate the latest energy-efficient technologies
The above considerations are of course only one part of the RAID framework, but offers a clear start to address the sustainability issue with Generative AI. With a clear focus on the sustainability dimension, businesses can strategically adopt Gen AI and still work towards sustainability targets.
"Businesses need to view AI not just as a technological advancement but as an opportunity to lead in sustainability. By adopting energy-efficient models and optimizing data management, companies can significantly reduce their environmental impact while maintaining high performance and innovation standards."
- Frida Holzhausen, Management consultant at Algorithma
As data centers switch to AI there will be a significant increase in energy consumption. For an average data center in Sweden, with an annual consumption of 30 GWh of energy before investment in additional GPU capabilities, the net increase of energy consumption in 2024 will be approx 20 GWh with an assumption of 70 % YoY increase.
Adopting a RAID approach can lead to substantial financial and environmental benefits. Assuming an average electricity price ranging from 0,8 to 1,2 SEK per kWh, and by implementing RAID principles: energy-efficient components, modern storage solutions, optimized server utilization, a sound deployment strategy, AI architecture, model design, and continuous AI operations, a conservative estimate suggests a 25% reduction in energy consumption. This reduction on the additional 20 GWh from AI use, would result in annual energy savings of 5 GWh. At an average electricity price of 1 SEK per kWh, this translates to cost savings of approximately 5 MSEK per year (ranging from 4-6 MSEK, depending on specific electricity costs).
Furthermore, this energy efficiency offset not only reduces costs but also has a substantial environmental impact. Given Sweden's low carbon emission factor of 0.013 metric tons CO2 per MWh, the annual reduction in energy consumption would lead to a decrease in carbon emissions by approximately 65 metric tons of CO2 (note: in Germany this would be approx. 25-28 times bigger due to the energy mix). In the Swedish context, the offset corresponds to 500 000 people flying one-way between Stockholm and Gothenburg. These quantifiable benefits highlight the value of RAID in enhancing sustainability while maintaining high performance and operational efficiency in data centers.
However, it is important to consider the broader energy market dynamics. The surge in demand for green energy driven by data centers and other high-energy-consuming sectors may outstrip Sweden's renewable energy supply, potentially leading to the import of less clean energy from neighboring regions. At the same time, it could reduce Sweden's energy exports to other countries, potentially causing them to rely more on non-renewable energy sources. This shift could negatively impact the overall carbon footprint of Sweden and its neighbouring countries, and counteract some of the environmental benefits achieved through energy efficiency measures.
Sweden is known for its reliable and sustainable energy, making it an attractive location for energy-intensive AI infrastructure which has also been highlighted e.g. through high-profile announcements. This raises important questions about energy demands and the potential need for substantial investments in nuclear power and renewable energy. Companies with significant compute needs have even indicated a willingness to invest in small modular nuclear reactors to meet the rising energy demands. This reflects a broader strategy where investments in renewable energy goes hand in hand with expansion of data center infrastructure.
By focusing on energy optimization, data centers can align better with sustainability goals, meet regulatory requirements, and improve overall cost-effectiveness. This strategic approach supports both environmental responsibility and economic viability, making RAID a compelling framework for modern data center management. However, stakeholders must also advocate for increased investment in domestic renewable energy capacity to mitigate the risk of relying on imported dirty energy.
Getting started with RAID sustainability
At Algorithma, we partner with Elastx to enable our clients to take a measured approach to Generative AI deployments.
Strategic approach to deployment; jointly we work through the use-cases assuring the right tech is used for the right application
Architecting for RAID; working through the setup to keep humans in the loop, making sure the computational tasks are scoped to the business problem
Model design and sizing, incl. HW selection offered as a service; building models that are fit for purpose, avoiding general models, to ensure quality application of tech and reduce computational needs
Right-sized and fit for purpose AI operations; continuously working with our own products and with the customer to prune and optimize models in production, incl. monitoring and optimization of loads
In addition, we work to optimize the energy consumption in the data center, all located strategically in Sweden, ensuring usage of green energy for operations. Through this approach we can offer our clients an easy way to deploy Generative AI.
Follow us for more updates on this topic as we continue to explore responsible AI by design.
