What sets AI and its current applications apart from other iterations of computing?
Jariwala: It’s a totally new paradigm in terms of function. Think back to the very first computer, the Electrical Numerical Integrator and Computer (ENIAC) we have here at Penn. It was built to do math that would take too long for humans to calculate by hand and was mostly used for calculating ballistics trajectories, so it had an underlying logic that was straightforward: addition, subtraction, multiplication, and division of, say, 10-digit numbers that were manually input.
Lee: Computing for AI has three main pieces. One is data pre-processing, which means organizing a large dataset before you can do anything with it. This may involve labeling the data or cleaning it up, but basically you’re just trying to create some structure in it.
Once preprocessed, you can start to ‘train’ the AI; this is like teaching it how to interpret the data. Next, we can do what we call AI inference, which is running the model in response to user queries.
Jariwala: With AI, it’s less about crunching raw numbers and more about using complex algorithms and machine learning to train and adapt it to new information or situations. It goes beyond manually entering a value, as it can draw information from larger datasets, like the internet.
This ability to gather data from different places, use probabilistic models to weigh relevance to the task at hand, integrate that information, and then provide an output that uncannily resembles that of a human in many instances is what sets it apart from traditional computing. Large language models, like ChatGPT, showcase this new set of operations when you ask it a question and it cobbles together a specific answer. It takes the basic premise of a search engine but kicks it up a gear.
What concerns do you have about these changes to the nature of computation?
Lee: As AI products like ChatGPT and Bing become more popular, the nature of computing is becoming more inference based. This is a slight departure from the machine-learning models that were popular a few years ago, like the DeepMind’s AlphaGO—the machine trained to be the best Go player—where the herculean effort was training the model and eventually demonstrating a novel capability. Now, massive AI models are being embedded into day-to-day operations like running a search, and that comes with trade-offs.
What are the material and resource costs associated with AI?
Jariwala: We take it for granted, but all the tasks our machines perform are transactions between memory and processors, and each of these transactions requires energy. As these tasks become more elaborate and data-intensive, two things begin to scale up exponentially: the need for more memory storage and the need for more energy.
Regarding memory, an estimate from the Semiconductor Research Corporation, a consortium of all the major semiconductor companies, posits that if we continue to scale data at this rate, which is stored on memory made from silicon, we will outpace the global amount of silicon produced every year. So, pretty soon we will hit a wall where our silicon supply chains won’t be able to keep up with the amount of data being generated.
Couple this with the fact that our computers currently consume roughly 20-25% of the global energy supply, and we see another cause for concern. If we continue at this rate, by 2040 all the power we produce will be needed just for computing, further exacerbating the current energy crisis.
Lee: There is also concern about the operational carbon emissions from computation. So even before products like ChatGPT started getting a lot of attention, the rise of AI led to significant growth in data centers, facilities dedicated to housing IT infrastructure for data processing, management, and storage.
And companies like Amazon, Google, and Meta have been building more and more of these massive facilities all over the country. In fact, data center power and carbon emissions associated with data centers doubled between 2017 and 2020. Each facility consumes in the order of 20 megawatts up to 40 megawatts of power, and most of the time data centers are running at 100% utilization, meaning all the processors are being kept busy with some work. So, a 20-megawatt facility probably draws 20 megawatts fairly consistently—enough to power roughly 16,000 households—computing as much as it can to amortize the costs of the data center, its servers, and power delivery systems.
And then there’s the embodied carbon footprint, which is associated with construction and manufacturing. This hearkens back to building new semiconductor foundries and packaging all the chips we’ll need to produce to keep up with increasing compute demand. These processes in and of themselves are extremely energy-intensive, expensive and have a carbon impact at each step.
This interview was conducted by Nathi Magubane. Read more about “The Hidden Costs of AI” at Penn Today.