Gemini's environmental footprint and efficiency

Our approach to energy innovation and the environmental footprint of AI

The AI represents one of the most important technological transformations of our time, which will become more evident in the next decade. Applied in areas such as medicine, energy, autonomous systems and quantum computing, the AI is poised to help people address important societal challenges, whether it's helping students learn, diagnosing cancer early, improving the safety of complex transportation and cybersecurity systems, or even predicting the path of forest fires for the rescuers.

See more articles from iGuRu.gr when you search for news on Google.

iGuRu.gr preferred source

Harnessing the power of AI will require strong energy infrastructure, more efficient energy use, and new innovative technology solutions. We are approaching this issue from many angles, investing in new infrastructure, designing smarter and more resilient grids, and expanding both established and new , innovative clean energy sources. At the same time, we are also focused on maximizing efficiency at every level of our operations, from the design of our purpose-built equipment to the software and models running in our data centers.

In order to improve the energy efficiency of AI , a clear and comprehensive understanding of the environmental footprint of AI is important . To date, detailed data on the energy and environmental impact of AI inference has been limited.

Today, we’re helping to bridge that gap by releasing a comprehensive methodology for measuring the energy, water, and carbon footprint of Google ’s AI models . It’s important to note that our work on model efficiency is progressing rapidly. Over a 12-month period, while delivering higher-quality answers, the average energy consumption and carbon footprint per text prompt in Gemini Apps decreased by 33x and 44x, respectively. Based on our recent analysis, we found that our efficiency work is proving effective, with the energy consumed for the average user request being equivalent to watching less than nine seconds of TV. These developments build on our long-standing commitment to data center efficiency. In 2024, for example, we reduced data center energy emissions by 12%, even though electricity consumption increased by 27% year-over-year, due to the expansion of our business and services.

As we continue to invest in the technology and innovation needed to meet significant new energy demands, transparency is key to progress. We hope this study will contribute to ongoing efforts to develop efficient AI at this critical time for energy, sustainability, and scientific discovery, for the benefit of all. Read more about how we arrived at our calculations and our full-stack approach to energy innovation, and delve deeper into our white paper.

Author : Ben Gomes​​

How much energy does Google AI use ? We did the calculations

Editors :​

1. Amin Vahdat , Vice President/General Manager, Vice President/General Manager, AI and Infrastructure, Google Cloud

2. Jeff Dean , chief scientist, Google DeepMind and Google Research

AI is unlocking scientific discoveries, improving healthcare and education, and could add trillions to the global economy. Understanding the environmental footprint of AI is critical, but detailed data on the energy and environmental impact of AI inference , that is, using a trained AI model to make predictions or generate text or images, has been limited. As more people use AI systems , the importance of inference efficiency increases.

This is why we are releasing a technique studywhich details our overall methodology for measuring the impacts of exhortationsof Gemini: in energy, emissions and water. Using this methodology, we estimate that the average text prompt of the Apps Gemini: uses 0,24 watt-hours (Wh) of energy, emits 0,03 grams of carbon dioxide equivalent (gCO₂e) and consumes 0,26 milliliters (or about five drops) of water¹These numbers are substantially lower than many public calculations. The energy impact per promptis equivalent to watching television for less than nine seconds.

At the same time, our AI systems are becoming more efficient through research innovations and improvements in software and hardware performance. For example, over a recent 12-month period, the energy and overall carbon footprint of the average Gemini Apps text prompt were reduced by 33x and 44x, respectively, ² while delivering higher-quality answers. These results are based on our most recent data center energy emissions reductions and our work to promote carbon-free energy and water replenishment. While we’re proud of the innovation behind our efficiency gains to date, we’re committed to continuing to make meaningful improvements. Here’s a closer look at these ongoing efforts.

Calculating the environmental footprint of AI At Google

Detailed measurement allows us to make comparisons between different AI models and the equipment and energy they use, while also enabling efficiency optimization at the system level, from equipment and data centers to the models themselves. By sharing our methodology, we hope to increase consistency in calculating resource consumption and AI efficiency across the industry.

Measuring the footprint of AI-powered tasks is not simple. We developed a comprehensive approach that takes into account the real-world conditions of AI operations at Google scale , which include:

• Full system power potential: This includes not only the energy and water used by the main AI model during active computing, but also the actual utilization of processing power at production scale, which can be much lower than theoretical maximums.

• Idle machines: To ensure high availability and reliability, production systems require a degree of provisioned capacity that is idle, but ready to handle traffic surges or failover in the event of a failure at any time. The energy consumed by these idle chips must be factored into the overall energy footprint.

• CPU and RAM: Model execution AI is not done exclusively on machine learning accelerators, such as TPU and GPU. The CPU and RAM of the host computer also play a crucial role in providing AI and they use energy.

• Indirect data center costs: The energy consumed by the IT equipment running AI workloads is only part of the story. The infrastructure that supports these computations, namely cooling systems, power distribution, and other indirect data center costs, also consumes energy. Auxiliary energy efficiency is measured by a metric called Power Usage Effectiveness ( PUE ).

• Data Center Water Consumption : To reduce energy consumption and related emissions , data centers often consume water for cooling. As we optimize our AI systems to be more energy efficient, their overall water consumption naturally decreases as well.

Many current energy consumption calculations AI include only the active consumption of machines, ignoring several of the critical factors discussed above. As a result, they represent theoretical efficiency rather than actual operational efficiency at scale. When we apply this incomplete methodology that only considers active consumption TPU and GPU, we estimate that the waisttext prompt Gemini: uses 0,10 Wh energy, emits 0,02 gCO₂e and consumes 0,12 mL water. The best that can be said about this is that it is an optimistic scenario, which however greatly underestimates the actual operational footprint. AI.

Our comprehensive methodology estimates (0,24 Wh energy, 0,03 _gCO2e , 0,26 mL water) take into account all critical elements of global AI service provision . We believe this is the most comprehensive view of the overall AI footprint.

Our comprehensive approach to AI and efficiency AI

Gemini ’s impressive efficiency benefits come from Google ’s full-stack approach to AI development , which includes custom hardware and highly efficient models, as well as the robust delivery systems that enable the use of those models. We’ve built efficiency into every layer of AI , including:

• More efficient model architectures: Gemini models are based on the Transformer model architecture developed by Google researchers and provide a 10-100x efficiency increase over previous state-of-the-art architectures for language modeling. We design models with inherently efficient structures, such as Mixture - of - Experts ( MoE ) and hybrid reasoning . MoE models , for example, allow us to activate a small subset of a large model that is specifically needed to answer a query, reducing computation and data transfer by 10-100x.

• Efficient algorithms and quantization: We continuously improve the algorithms that support our models with methods such as Accurate Quantized Training ( AQT ) , to maximize efficiency and reduce energy consumption for service provision, without compromising on the quality of responses.

• Optimized inference and service delivery: We are constantly improving our model offering AI for greater responsiveness and efficiency. Technologies such as visual decoding serve more answers with fewer chips, allowing a smaller model to make predictions that are quickly verified by a larger model. This is more efficient than having the larger model make many successive predictions on its own. Techniques such as distillation create smaller, more efficient models ( Gemini: Flash and Flash Small) to provide services that use our larger, more capable models as teachers. Faster hardware and machine learning models allow us to use larger batch sizes when handling requests, without sacrificing latency targets.

• Customized equipment: We design the TPU from scratch for over a decade, to maximize performance per wattWe also design the models together. AI and TPU, ensuring that our software takes full advantage of our hardware and that our hardware can effectively run future software AI ours, when both are ready. The TPU latest generation, ironwood, is 30 times more energy efficient than our first TPU which was available to the public and much more energy efficient than the CPU general use for inference.

• Optimized Idleness: Our serving stack utilizes CPUs extremely efficiently and minimizes TPU idleness , dynamically moving models based on demand in near real-time, rather than using a “ set it and forget it ” approach.

• Machine learning software stack: The XLA ML compiler, Pallas cores , and Pathways systems enable model computations expressed in higher-level systems, such as JAX , to run efficiently on TPU hardware.

• Highly efficient data centers: Google 's data centers are among the most efficient in the industry, with an average PUE of 1,09 across all equipment.

• Responsible data center operations: We continue to add clean energy generation as part of our ambition for 24/7 carbon-free operations , while advancing our goal of replenishing 120% of the freshwater we consume on average in our offices and data centers. We are also optimizing our cooling systems, balancing the local trade-off between energy, water and emissions, conducting science-based hydrocrit health assessments to guide cooling type selection and limiting water use in high-stress locations.

Our commitment to efficiency AI

Gemini ’s efficiency gains are the result of years of work, but they’re just the beginning. Recognizing that demand for AI is growing, we’re investing significantly in reducing the cost of power and water required per prompt . By sharing our findings and methodology, we aim to drive progress across the industry towards more efficient AI . This is essential for the responsible development of AI.

guest

Guest Post: I saw openly and entered!