Halving the GPU carbon footprint

With all the talk of AI datacentres taking as much power as small cities, our researchers developed a technology that can halve the power consumption of GPUs in completing their workload. Wow!

In most GPU-intensive situations, like for LLM training, inference etc. the program runs on both the CPU and GPU. Often the GPU is only performing calculations for a fraction of the time, but is consuming power constantly. We can dynamically reassign and redirect running programs to maximise the utilisation of the GPU so that it is performing valuable calculations for almost the entire time it is powered on. [Link for more information.]

We thought that this would have a massive impact on global power consumption because all GPU users would want to reduce their power usage if it was as easy as adding a middleware component. However the companies we have approached have had a variety of reasons not to want to optimise GPU utilisation.

We feel certain that there must be segments that would jump at the cost and carbon saving benefits of the technology and are eager to create open-innovation partnerships and collaborations with GPU market and technology experts to explore this. Do you have the insight we need or additional technology to turn this into a solution?