Why GPGPU

Decentralized computing services were initially used for large-scale data storage and processing, web data collection, and simulations, but recently they have been utilized in more general cases. For example, simple Dapps run on the Ethereum blockchain network, or independent private blockchains are constructed to provide B2C services.

This backdrop is due to the fact that decentralized computing services are more efficient and streamline processes compared to traditional centralized systems. They are also more cost-effective. Dapps running on Ethereum, for instance, require minimal management resources if the contracts are well-written. Unlike centralized services, there is no need to constantly check server status or security. When it comes to overcoming the physical limitations of running computer-based services, decentralized computing offers the most suitable solution.

Decentralized computing services can effectively address the challenges currently faced by the AI and machine learning industries. While it took 30 years from the introduction of machine learning concepts in the 1990s to the emergence of OpenAI's ChatGPT, the sophistication of ChatGPT is now increasing tenfold each year, leading to explosive advancements. According to OpenAI, as performance increases, the computing resources required to achieve optimal machine learning results are also increasing by more than fivefold annually.

The rapid advancement of AI technology has led to a surge in demand for GPU resources, resulting in significantly higher costs for obtaining these resources in the market. This trend hampers AI development and imposes high entry costs on teams attempting to break into the market. While AI technology continues to rise exponentially, the growth in hardware capabilities is lagging, causing numerous issues.

Decentralized computing services offer a compelling solution by providing a more efficient, cost-effective, and scalable infrastructure to support the growing needs of AI and machine learning industries.

Increasing Complexity of AI Models

• Size and Complexity of AI Models: Recent AI models, especially large language models (LLMs) and multimodal models, are becoming increasingly complex. For example, models like OpenAI's GPT-4 and Google's Gemini Ultra require substantial computational resources, costing $78M and $191M, respectively​​.

• Need for High-Performance Computational Resources: These large-scale models require massive datasets and complex computations to achieve high accuracy, leading to increased demand for high-performance GPUs. Since 2023, the costs associated with training AI models have grown exponentially, highlighting the critical importance of GPUs in AI research​​.

Increased Use of AI in Various Applications

• Autonomous Driving and Healthcare: AI models are being utilized in highly computation-intensive applications such as autonomous driving and precision medical diagnostics, resulting in an explosive increase in GPU demand. These models require real-time data processing and analysis, making high-performance GPUs essential​​.

• Industry and Corporate Applications: Many companies are adopting AI to enhance productivity and improve quality, which is another major driver of GPU demand. For instance, companies are using AI for customer service automation and personalized content delivery, which also necessitates high-performance computational resources​​.

Increased Use of AI in Education and Academic Research

• AI Research in Educational Institutions: Universities and research institutions are fostering future talent through AI-related research and educational programs, which require substantial GPU resources. Students and researchers learning and studying AI technologies need high-performance computational resources, significantly driving GPU demand​​.

• Increased Industry-Academia Collaboration: Recently, collaboration between industry and academia has been increasing, broadening the scope and depth of AI research. In 2023, 21 notable AI models were developed through industry-academia collaboration​​.

Due to these factors, the spread of AI research and development is drastically increasing GPU demand, potentially leading to future GPU supply shortages. Therefore, GPU manufacturers need to expand production capabilities and address supply chain issues to meet this growing demand.

Mission

The problem is that for the development of machine learning, deep learning, and similar technologies, diverse and repetitive training and simulations are essential. The AI industry is no longer a story of the future; it is our reality. What used to take 30 years is now achieved in just one year, continuing to drive dazzling innovation.

However, if the resources necessary for industrial growth are not adequately supplied, the growth rate will slow down and reach a limit. A gap between the development speeds of software and hardware can lead to not only a decrease in growth but also bottlenecks, causing significant disruptions.

GPGPU can innovatively solve this problem by leveraging the idle GPU resources of companies and individuals worldwide. Most GPUs owned by individuals and companies remain idle for the majority of the time. Through GPGPU, GPU resources can be managed efficiently, overcoming cost and physical limitations, and contributing to accelerating the life-changing innovations that AI will bring.

Using multiple NVIDIA RTX 4090s in AI development offers high cost efficiency and excellent performance, making it a suitable choice for small to medium-sized AI startups. While high-performance GPUs like the A100 might be a better choice for specific scenarios such as large-scale data centers, the RTX 4090 is a highly competitive option considering budget and performance.

Nonetheless, traditional cloud services only offer industrial devices like the A100 and H100, which are more effective in terms of power consumption efficiency. Since GPGPU is a decentralized GPU cloud service, it can provide more diverse options by alleviating power consumption concerns, offering users greater flexibility and choice.

Economic Benefits for GPU Providers

A decentralized GPU cloud service offers significant advantages not only for GPU resource users but also for providers. Building a GPU server center involves substantial costs, which are typically covered by the company’s revenue model. However, by providing GPU resources through a decentralized GPU cloud like GPGPU, additional revenue streams can be generated.

Examining the scale of GPU server centers operated by global companies, we find that various entities such as research institutions, game developers, and film production companies run GPU clusters of different sizes.

Game Developers:

• Electronic Arts (EA): EA uses high-performance GPU servers for real-time graphics rendering and game engine development. They leverage NVIDIA’s RTX servers to implement complex graphic processing and AI-based game elements.

• Ubisoft: Ubisoft employs large GPU clusters to develop its various game titles. These clusters support game development, testing, real-time rendering, and AI model training.

• Epic Games: Epic Games operates high-performance GPU server centers for the development and optimization of Unreal Engine. They use NVIDIA’s latest GPUs to maximize game graphics and physics engine performance.

Film Production Companies:

• Weta Digital: Weta Digital utilizes large GPU clusters to create visual effects (VFX) for blockbuster films like "The Lord of the Rings." These clusters, comprising thousands of GPUs, render high-quality graphics and complex visual effects in real-time.

• Pixar Animation Studios: Pixar uses large-scale GPU servers for animation film production. These servers accelerate complex 3D rendering and visual effects tasks, reducing production time and enabling higher-quality graphics.

• Industrial Light & Magic (ILM): ILM operates high-performance GPU clusters for producing special effects in various films. They use the latest GPUs for real-time rendering and visual effects production, which is crucial for the complex graphics work required in blockbuster movies.

Game developers and film production companies operate high-performance GPU server centers to support complex graphic processing, AI model training, and real-time rendering. GPGPU targets these entities as potential GPU providers, including both individuals and institutions that own GPU server centers.

These GPU server centers often have significant idle times. On average, GPU servers maintain about 50% utilization, remaining idle outside of specific tasks or periods of intensive use, such as rendering and simulation work for game developers or film production schedules.

Through GPGPU, rental agreements can be established on a periodic basis, or a dynamic allocation system can be implemented to rent out GPUs only when they are idle. A decentralized GPU cloud service can enhance cost efficiency for both users and providers.

By leveraging the underutilized GPU resources of game developers and film production companies, GPGPU offers a cost-effective solution, benefiting all parties involved. This approach not only maximizes the use of existing hardware but also generates additional revenue for GPU providers, promoting a more efficient and sustainable AI and computing ecosystem.