Acceleration through AI

The AI and ML Workload Challenge

AI and ML technologies have become indispensable across a spectrum of industries, from healthcare to finance and manufacturing to entertainment. AI’s advancement necessitates considerable computational power, with the energy consumption for training a single AI model exceeding that of 100 homes in a year. This substantial energy use, highlighted by Google’s report that machine learning accounted for about 15% of its total energy consumption over three years, underscores the critical need for prioritizing the development of more efficient computing infrastructures and algorithms to mitigate the environmental impact and ensure AI contributes positively to the energy sector’s efficiency and sustainability goals.  

An Innovative Cooling Solution

In response to these challenges, data centers have been exploring unconventional methods to achieve optimal performance. One such pioneering technique involves submerging GPUs in non-conductive oil. This method offers several advantages over traditional cooling approaches. 

1. Higher Thermal Conductivity:

Non-conductive oils have a higher thermal conductivity than air, allowing for more efficient heat transfer away from the GPU. This means that the cooling system can operate more efficiently, reducing the amount of energy required to maintain optimal operating temperatures. 

2. Reduced Cooling Infrastructure:

Traditional air-cooling systems often require a combination of heatsinks, fans, and sometimes even air conditioning units to manage the heat generated by high-performance GPUs. Each of these components consumes electricity. In contrast, an oil immersion system can often achieve the same or better cooling performance with less complex and energy-intensive infrastructure. 

3. Elimination of Fans on GPUs:

When GPUs are submerged in non-conductive oil, the fans typically mounted on the GPUs can be removed. This directly reduces the power consumption of each GPU, as the fans themselves can consume a significant amount of electricity, especially in large-scale deployments. 

Quantitatively, energy savings can vary widely but are often reported to be in the range of 10% to 50% or more, depending on the specifics of the application and the efficiency of the cooling system design. 

The Environmental Impact

In addition to the energy savings, what sets this particular HPC environment further apart is its commitment to sustainability. Instead of allowing the residual heat generated by submerged GPUs to go to waste, a solution has been devised. This surplus heat is redirected to a local city heating system, providing warmth to thousands of households in the community. This integration of technology and sustainability reflects a dedication to reduce the carbon footprint and make a positive impact within the local community. 

The Future of Sustainable AI

This innovative HPC environment serves as a prime example of how sustainability and speed can coexist in the world of AI and ML workloads. In fact, the future is here, right now: through immersed cooling solutions by Asperitas, companies like Bytesnet lead the way and have this solution implemented as we speak. This approach exemplifies the possibility of a world where cutting-edge technology thrives without compromising our commitment to the planet. Being “sustainable and yet fast” is not only possible but essential in the 21st century, highlighting the potential for a more sustainable future.