<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Vineeth Venukumaran Nair on Medium]]></title>
        <description><![CDATA[Stories by Vineeth Venukumaran Nair on Medium]]></description>
        <link>https://medium.com/@vineethv31?source=rss-45759c373167------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*kfZdpJQ4L6ETJcUn</url>
            <title>Stories by Vineeth Venukumaran Nair on Medium</title>
            <link>https://medium.com/@vineethv31?source=rss-45759c373167------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 08:59:16 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@vineethv31/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[AI’s Carbon Footprint and Strategies for Optimization]]></title>
            <link>https://medium.com/@vineethv31/ais-carbon-footprint-and-strategies-for-optimization-c49a2f9ad2bf?source=rss-45759c373167------2</link>
            <guid isPermaLink="false">https://medium.com/p/c49a2f9ad2bf</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[carbon-footprint]]></category>
            <category><![CDATA[mainframe-modernization]]></category>
            <dc:creator><![CDATA[Vineeth Venukumaran Nair]]></dc:creator>
            <pubDate>Sun, 15 Dec 2024 10:41:35 GMT</pubDate>
            <atom:updated>2024-12-15T10:41:35.675Z</atom:updated>
            <content:encoded><![CDATA[<blockquote><strong>“A single ChatGPT query uses 2.9 watt-hours of electricity, which is nearly 10 times more than a Google search”</strong></blockquote><blockquote><strong>“Training a model like Generative Pre-trained Transformer 3 (GPT-3) uses roughly 1,300 megawatt hours (MWh) of electricity, which is similar to the annual power consumption of 130 homes in the US”</strong></blockquote><p>Artificial Intelligence (AI) has become a transformative force across various industries, but its rapid growth comes with a significant environmental cost. The carbon footprint of AI is a growing concern, as the energy required to train and run AI models is substantial. This article explores the carbon footprint of AI and discusses strategies for optimizing AI to reduce its environmental impact.</p><p><strong>AI’s Carbon Footprint</strong></p><p>AI models, especially large language models (LLMs), require vast amounts of energy to train and run. Training a single model can produce significant carbon dioxide emissions. For instance<strong> </strong>training a model like Generative Pre-trained Transformer 3 (GPT-3) uses roughly 1,300 megawatt hours (MWh) of electricity. The energy consumption doesn’t stop at training; running these models also contributes to ongoing emissions.</p><p>The carbon footprint of AI is influenced by several factors, including the efficiency of code written, energy source used to power data centers and the efficiency of the hardware. Data centers in regions with cleaner energy grids, such as those powered by nuclear or renewable energy, have a lower carbon footprint compared to those in regions reliant on fossil fuels.</p><p><strong>Strategies for Optimization</strong></p><p>To mitigate the environmental impact of AI, several strategies can be employed:</p><p><strong>1. Hyperparameter Tuning</strong></p><p>Optimizing the parameters of AI models can improve efficiency and reduce energy consumption. This involves fine-tuning the settings that govern the learning process. Some examples are using —</p><p><strong>a. Learning Rate — </strong>Controls how much to change the model in response to the estimated error each time the model weights are updated. Setting a too high learning rate can cause the model to converge too quickly to a suboptimal solution. Setting it too low can make the training process unnecessarily long.</p><pre># Define the optimizer with a specific learning rate <br>optimizer = Adam(learning_rate=0.001)</pre><p><strong>b. Number of Epochs</strong>: The number of times the learning algorithm will work through the entire training dataset. Too few epochs might result in underfitting, while too many could lead to overfitting. Finding the right balance is key.</p><pre>• # Train the model and set the number of epochs <br>model.fit(x_train, y_train, epochs=50, batch_size=32, validation_data=(x_val, y_val))</pre><p><strong>c. Dropout Rate</strong>: A regularization technique where randomly selected neurons are ignored during training. Helps in preventing overfitting. A higher dropout rate means more regularization.</p><pre>• # Define the model <br>model = Sequential([Dense(128,activation=&#39;relu&#39;,input_shape=(input_dim,)),Dropout(0.5), # 50% dropout rate</pre><p><strong>d. Optimizer Choice</strong>: Algorithms used to change the attributes of the neural network such as weights and learning rate to reduce the losses. Choosing between optimizers like SGD, Adam, RMSProp, etc., can have a significant impact on training efficiency and model performance.</p><p>2. <strong>Quantization</strong></p><p>Quantization in machine learning is a technique where the precision of the weights and activations of a model is reduced, typically from 32-bit floating point to 8-bit integers. This can significantly reduce the computational load and energy consumption without a substantial loss in model accuracy. Below is an example of how you can perform quantization using TensorFlow and its tf.quantization module.</p><pre># Quantization: Convert the model to a quantized version <br>converter = tf.lite.TFLiteConverter.from_keras_model(model) <br>converter.optimizations = [tf.lite.Optimize.DEFAULT] <br>quantized_model = converter.convert()</pre><p>3. <strong>Smaller models</strong></p><p>All other things being equal, a smaller model will have fewer layers with fewer filters and will therefore require less computing and less memory. As we’ll see, there are ideas that make a model smaller on a tactical basis, but if <em>MobileNet</em> were sufficient for everything, then we wouldn’t need <em>YOLO</em>. Instead, the smaller models are chosen when the computing platform and energy source so require, with accuracy as a potential tradeoff.</p><p>4. <strong>Using Specialized Hardware</strong></p><p>Using specialized hardware like GPUs (Graphics Processing Units) and CUDA (Compute Unified Device Architecture) can significantly accelerate computation-intensive tasks and reduce energy consumption.</p><p>GPUs are particularly well-suited for parallel processing tasks such as deep learning. Here’s an example using TensorFlow to train a neural network on a GPU</p><pre># Check if GPU is available<br>print(&quot;Num GPUs Available: &quot;, len(tf.config.experimental.list_physical_devices(&#39;GPU&#39;)))<br><br># Train the model on GPU<br>with tf.device(&#39;/GPU:0&#39;):<br>  model.fit(x_train, y_train, epochs=10, validation_data=(x_test, y_test))<br></pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/501/1*qUFuqCDfg781iMwZAKww1g.png" /></figure><p>Additionally, IBM is also extending its Mainframe capability to support AI workloads. IBM Spyre is an AI accelerator chip designed to enhance the AI capabilities of IBM Z mainframe systems. It was previewed at the Hot Chips 2024 conference and is expected to be available in 2025.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/510/1*WuLVrnxxBoTJVjO04y1DLA.png" /></figure><p><a href="https://research.ibm.com/blog/spyre-for-z">Introducing the IBM Spyre AI accelerator chip — IBM Research</a></p><h3>Reference —</h3><p><a href="https://www.weforum.org/stories/2024/07/generative-ai-energy-emissions/#:~:text=As%20one%20AI%20researcher%20said,doubling%20roughly%20every%20100%20days.&amp;text=How%20is%20the%20World%20Economic%20Forum%20creating%20guardrails%20for%20Artificial%20Intelligence?&amp;text=In%20response%20to%20the%20uncertainties,transparent%20and%20inclusive%20AI%20systems."><em>AI and energy: Will AI reduce emissions or increase demand? | World Economic Forum</em></a></p><p><a href="https://semiengineering.com/11-ways-to-reduce-ai-energy-consumption/"><em>11 Ways To Reduce AI Energy Consumption</em></a></p><p><a href="http://www.github.com/"><em>www.github.com</em></a></p><p><a href="https://developer.nvidia.com/"><em>NVIDIA Developer</em></a></p><p><a href="https://www.ibm.com/blog/announcement/telum-ii/"><em>New Telum II Processor and IBM Spyre Accelerator: Expanding AI on IBM Z — IBM Blog</em></a></p><p><em>GenAI’s Carbon Footprint: A New Challenge For Corporations</em></p><p><em>AI is poised to drive 160% increase in data center power demand | Goldman Sachs</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c49a2f9ad2bf" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>