With the rising popularity of generative artificial intelligence (genAI), the so-called hyperscalers (Microsoft, Google, AWS, etc) are racing to build new data centers. These data centers consume a lot of energy, consume a lot of water for cooling and emit a lot of CO2. This naturally raises concerns among users. 

Here we give a brief FAQ to outline how Raffle’s search and generative AI services are run, how much energy and water they consume and how much energy generative AI uses relative to old fashioned internet search and our total everyday electricity consumption.

Summary: Raffle’s servers, located in Denmark and Sweden, use good industry practices with low energy and water consumption and low CO2 emissions due to use of renewable energy resources. 

What hardware does Raffle run on?

You have to distinguish between serving and model training/research. When users interact with Raffle search, summary and chat, they are using the served Raffle models. For reliability, scaling and security reasons the serving is done in the cloud. Raffle uses a Microsoft Azure datacenter located in Sweden. We train and develop our models on our own computer which is located in a datacenter in Denmark.

What is the energy and water consumption of Raffle serving in the Azure datacenter in Sweden?

According to Azure, their datacenter Sweden will be CO2 and water consumption neutral by 2030. This is achieved by using renewable energy sources and collecting rainwater. Their 2022 PUE (power usage effectiveness) is 1.172. Low PUE is better. A value of 1 means that all energy used goes towards computation. Their 2022 water consumption is 0.16 L/kWh (liters per kWh electricity used). Below, we will approximately translate these numbers to consumption per question.   

What is the energy and water consumption of Raffle model training and research in the Digital Reality/Interxion datacenter in Denmark?

According to the Digital Reality factsheet, the datacenter uses “advanced cooling technologies to reduce overall energy consumption” and “utilizes 100% renewable energy sources to minimize carbon footprint”. More specifically, cooling is performed by free cooling (cool air and circulated water without energy intensive compression) and compressor cooling. Interxion has made an agreement with Vattenfall for 24/7 delivery of renewable energy. 
The most criticized cooling method is water evaporation using drinking water. It is energy efficient but can put strain on scarce water resources in some places. Water evaporation is also used in some datacenters in Denmark, but not in Raffle’s datacenter in Denmark. It can be expected that the EU will put caps on this use in the coming years. 

How does the footprint of generative AI compare to that of other things we do in our everyday life?

This is a difficult question because the answer will depend upon exactly how sustainable the datacenters are. But it is possible to give some approximate numbers. A genAI answer uses approximately 3Wh of electricity which should be compared with an old-fashioned Google internet search that uses one-tenth of that or approximately 0.3Wh. These numbers can be put into perspective by comparing our average everyday energy and electricity consumption. Let us imagine a genAI power user who asks 100 questions per day. This amounts to 300Wh. A Dane uses on average 4300Wh per day for personal electricity consumption. The power user would then use approximately 6.25% of the entire electricity consumption on genAI. So if we all become genAI power users, we will get a substantial increase in our overall electricity consumption that can only be offset by making genAI more energy efficient and the grid less CO2 emitting. What is the water consumption of a genAI question?

Using the water consumption number from the Azure Sweden datacenter, this gives 3*0.16/1000 L = 0.00048 L of water for one query. Or said in a different way, asking Raffle Summary 1/0.00048 = 2083 questions will consume 1 liter of water!