ChatGTP’s sudden and arguably premature success has yet again exposed the usually overlooked link between so-called “virtual” digital technologies and very tangible infrastructure. Indeed, early adopters of the latest incarnation of GTP-3-based bots directly experienced repeated network and login failures. That surely happened to me a few times back in December when I first started playing with the newly launched computational agent. However, its creators promptly addressed these by adding new infrastructural capacity with much-needed assistance from one of the Big Tech siblings, whose hefty financial support also helped carry the day.
However, less well-known to end users and the general public are the overall infrastructure requirements to create, train and deploy ChatGTP. As indicated in a previous post, we can learn about such requirements by embracing a comprehensive production process framework that looks at the creation, distribution, exchange, consumption and disposal of all ChatGPT components. The first step is to take a peek at the actual production of the agent, a quasi-new member of the always-growing Machine Learning (ML) family. Here we encounter yet again the three phases encompassing the ML production cycle, including design and data, training and testing, and deployment for public consumption (with further training and refining based on the latter, as needed). 
ChatGTP’s production information is unavailable from one single source; what is out there is undoubtedly far from complete. The best sources I found were a couple of research papers from 2020 and 2022 and OpenAI’s website announcing the launch of the sophisticated and game-changer large language model. While the latter is quite digestible, the two papers demand a bit more patience and attention to find what we are searching for. Do not be discouraged by the technical jargon both proudly display. By the way, it is also critical to be fully aware of the various projects and code versions developed by OpenIA in the process.
The 2020 paper provides key insights. It explains what data was used and how it was processed before being fed to the algorithms. Essentially, the project used five already existing Internet datasets, Wikipedia and Reddit included. Some were cleaned and further fined tuned before starting the training process. The total size of the final dataset is not mentioned but is probably part of the terabytes (10^12) team. And yet, it does not cover all the data that is available on the Internet, estimated to be around 100 zettabytes (10^21) for 2023, most of which has been indexed by dominant search engines. Regardless, the model was trained on the tokenized version of the data, which generated 500 billion tokens, of which only 300 billion were used in the training phase. Assuming that each token comprises 16 bits, we get 600 gigabytes. The final model had 174.6 billion parameters, thus requiring almost 350 gigabytes of additional memory and storage space. Simplifying a bit, we can conclude that the data collected, cleaned and compiled in the data phase of the project amounted to 30%, give and take, of the total Internet data. Running data processing and tokenization processes to handle such data demands vast energy resources – although we do not have precise numbers. Notwithstanding, we can provisionally conclude that ChatGPT had low scope 1 but relatively high scope 2 emissions in the design and data phase.
In any event, running such a humongous model clearly demands very powerful computers with access to vast amounts of memory and storage space. Such beasts do exist, of course, some directly owned by and available to Big Tech companies. But, in this case, Microsoft Azure provided the much-needed computing and Internet infrastructure, thanks to its hyperscale data centers – where land, water, geography and people play as significant a role as technology and energy consumption.
The 2020 paper tells us that eight different models were trained, starting with a “small” GTP model with 125 million parameters. The final GPT-3 model thus ended up with 1,400% more parameters. It required 3,640 petaflops (10^15 floating point operations per second) – day, or 314 zettaflops (10^21) per 24-hour day, which we can compute by multiplying the former by the total number of seconds in a day (86400). That is larger than the estimated number of bytes on the Internet. The paper also clarifies that these numbers do not include some calculations that amount to an additional ten percent of the total. Now, 10 percent of such huge numbers are still huge, so tag it on. Remember that we do not know how many runs per model were executed. So the big number could actually be much more significant.
Curiously, while the paper acknowledges that GTP-3 did require intensive computing resources it does not provide any information on energy consumption. For example, the idea of flops per watt has been around for a while now, and some supercomputer specifications provide such a measure. Moreover, open-source libraries geared towards estimating AI emissions are readily available and could be immediately used by programmers. Not here. However, the paper quickly reports that, once deployed for public consumption, the GPT-3 model will consume 0.4 kilowatts-hour for every 100 pages generated. Clearly, OpenAI is s not as open as one should expect. Instead, it does not want to address emissions or ecological footprints like most other companies in the field.
Summarizing, we can partially conclude that GPT-3 data and design, and training phases can trigger substantial GHG emissions. At the same time, its deployment for public use is much more effective emissions-wise – assuming rebound effects will not eventually pop-up.
Raúl
