GPT Party 3.0. Dmitry Trifonov: Artificial Intelligence Computing Platforms GPT Party 3.0, the largest Russian-language networking event dedicated to artificial intelligence, took place in Silicon Valley on March 9-10. More than 450 people gathered at Plug and Play to meet leading experts, entrepreneurs, and investors, discuss the latest trends in artificial intelligence, and gain practical knowledge. At GPT Party, Dmitry Trifonov, co-founder and CEO of FiarCompute, a vendor-neutral artificial intelligence computing platform, spoke on the topic of “AI compute platforms.” Dmitry talked about why computing platforms are needed for AI, why there is a high demand for these computing platforms now, and how companies are solving this problem. Dmitry Trifonov, Founder of FairCompute I specialize in AI development, so I’ll talk about how to develop and scale AI for a large number of users. I’ve been a programmer or a manager of programmers for many years, and for the last ten years I’ve been working on computer vision and AI model infrastructure at various companies. In this presentation, I’ll try to answer three questions: why do we need AI computing platforms, why is there such a high demand for these platforms, and how are companies solving this problem. The first thing to note is that we tend to associate AI with smart people inventing algorithms. But in reality, most of the breakthroughs in AI have been achieved by scaling computing power. The second important point is that implementing AI solutions in production is extremely labor-intensive. It takes a huge amount of work to optimize and adapt these solutions for users. Let’s look at some historical examples. It all started when Alex Kryzhevsky popularized deep learning. They took an existing model, known since the 90s, and trained it on a huge amount of data, which significantly improved the results compared to what was available before. This marked the beginning of the deep learning renaissance. Later, with the advent of models like ChatGPT, even more data and computing power were used, thanks to which Microsoft built supercomputers. They used fairly standard models at the time, but training on huge amounts of data and investing millions of dollars significantly improved the results. Modern models require even more computation and data to further improve. The second point is that artificial intelligence has been a part of our lives for a long time. We all know about the hype around ChatGPT or Sora. However, since 2017, Apple products have been using computer vision and text processing technologies based on machine learning. In 2017, I joined the team that worked on Animoji. It was one of the few teams actively using machine learning, and many thought that running such models on a phone was crazy. For our team, this was a big challenge, and many did not believe in the possibility of implementing our ideas, but it worked. Most of the AI-based features in Apple products are almost invisible. They create a natural and improved user experience: photos become better, communication during calls and video chats is clearer, and the background in images is blurred. However, the real problem arose with Apple Vision Pro, where the number of calculations and functions requiring AI increased tenfold. We didn’t even believe we could run such a system, because it constantly scans the environment, tracks eye movements, faces, gestures, and people. Apple Watch also uses AI — all made possible by specialized hardware the company has developed: more powerful processors optimized for AI tasks. The answer to the question of why the demand for computing power is growing can be divided into two main reasons. The first is that more and more industries are starting to use artificial intelligence, which automatically increases the demand for computing power. The second reason is that the complexity of the AI models themselves is growing very quickly, significantly outpacing the growth of available computing power on the planet. To give an example from the article: since 2010, during the heyday of deep learning, the rate at which computing power requirements have doubled has been approximately every five to seven months, while the doubling of available computing power occurs approximately every 24 months. The complexity of computations is growing due to changes in architecture and the introduction of new, more complex methods. For example, instead of viewing an image once, a neural network analyzes it approximately 50 times. This significantly increases the computational costs. There is also a transition to transformers and diffusion models, where the result is formed gradually, which also requires multiple iterations. In the future, with the increasing use of 3D data, audio and video, the demands on computing power will only increase. What do companies typically do to respond to the growing demand for computing power? There are two main approaches. The first is to develop specialized chips, as Apple, Google, and now Amazon do. Many startups, such as Groq and SambaNova, have also attracted attention for their innovations in this area. The second approach is to acquire large quantities of graphics cards. For example, Mark Zuckerberg bought up many Nvidia graphics cards. Microsoft has developed computing platforms for OpenAI. In terms of efficiency, general-purpose hardware, such as those from Microsoft and Nvidia, historically performs better because it can run a variety of tasks. This hardware is not as efficient as specialized hardware, but it is more adaptable due to constantly changing architectures, making specialized chips obsolete. For example, Apple constantly updates its hardware, releasing new chips every two years and adapting to changes in technology, from conventional neural networks to transformers and large language models (LLM). Software also plays a key role. Despite AMD’s excellent hardware, they have yet to realize its full potential due to software limitations. This could change with new devices like the Apple Vision Pro and others that require specialized hardware to work directly with the user, rather than in big data centers. Such changes will require even more specialized hardware. For large companies, investing a billion dollars in developing new hardware is possible, but what about small companies? The cost of computing becomes critical, especially when creating a scalable product. If the computing costs are low, you can offer the solution for free. If they are high, you have to switch to subscriptions, and this directly affects the company’s competitiveness, especially when it comes to a mass AI product. Now large companies are buying up video cards and building huge data centers. They control significant computing power that is difficult to use effectively due to fluctuations in supply and demand, which sometimes leads to excess capacity. My idea is to resell unused computing resources to small companies, thereby reducing their costs and at the same time optimizing costs for large players.
