When it comes to the hottest new technology topic in the tech world, nothing beats ChatGPT. The discussions it has sparked about whether artificial intelligence (AI) can replace human work are ongoing. Setting aside the controversies of legality and ethics, ChatGPT can indeed accurately answer most general knowledge questions posed by users. Now, you might wonder, what does ChatGPT rely on to acquire this "omniscient" power?
As a conversational AI, the full name of ChatGPT is Chat Generative Pre-trained Transformer. It was developed by OpenAI and was released in November 2022. ChatGPT utilizes a large-scale language model based on the GPT-3.5 architecture (the latest open version). It was trained using reinforcement learning on Microsoft Azure's supercomputers, and fine-tuned using a proximal policy optimization algorithm. The model has a staggering parameter count of 175 billion. In summary, the foundation of ChatGPT lies in its large-scale language model, and its core competitiveness lies in its computational power.
The significant demand for computational power by ChatGPT can be illustrated by a set of data. The training of GPT-3.5 utilized a specialized AI computing system built by Microsoft, consisting of a high-performance network cluster with 10,000 V100 GPUs. The total computational power consumed amounted to approximately 3640 PF-days, which means that if performing calculations at a rate of 10 quadrillion operations per second, it would take 3640 days to complete. Meanwhile, the computational requirements of ChatGPT continue to expand. The large-scale language model underwent three iterations, with the parameter count increasing from 117 million in GPT to 175 billion in GPT-3. The pretraining data volume also grew from 5GB to 45TB. It is worth noting that a single training run of GPT-3 alone incurred a cost of $4.6 million. In practical scenarios, each question posed to ChatGPT incurs a cost of a few cents in computation. Therefore, for OpenAI, continuously obtaining computational support and controlling the high costs of computation are crucial. Currently, ChatGPT has a strong partnership with Microsoft, and OpenAI has stated that Microsoft Azure will be the sole designated cloud computing provider for ChatGPT, both now and in the future. This reveals the clear investment logic of Microsoft, providing capital and computational power in the short term, while capitalizing on the expanding demand for computational resources in the long run. Microsoft is well aware of the business opportunities behind data computation.
In other words, even if someone has access to the complex code of this large-scale model, not everyone can afford to run it. Therefore, the success of ChatGPT is not solely attributed to the intricate algorithms but also relies on the support of cloud computing services. OpenAI's collaboration with Microsoft goes beyond financial support and encompasses technical optimizations, including resource configurations for computing, storage, databases, networking, and more. For ChatGPT, leveraging the characteristics of the cloud on Microsoft Azure is fundamental for achieving high-performance computing, data storage and processing, global availability, elastic resource management, and cost-effectiveness. For instance, in recent times when ChatGPT faced a surge in website traffic from around the world, Microsoft Azure automatically provided additional resources such as CPUs and memory to handle the increased workload. Conversely, when the traffic subsides, resource configurations can be scaled down to save costs. Additionally, ChatGPT doesn't need to establish its own data centers but can rent the required resources from Microsoft Azure's cloud computing services on a pay-as-you-go basis. This eliminates maintenance costs and maximizes cost-effectiveness.
The immense popularity of ChatGPT not only reflects breakthroughs in AI technology but also signifies the development trend of big data in various industries. Cloud-based data computing and resource leasing replacing traditional purchasing are the general direction. By dynamically scaling resources to handle massive amounts of data, businesses can achieve cost reduction and efficiency improvement. This aligns with the original design intention of PieCloudDB Database, which aims to empower enterprises with cost-effectiveness through the utilization of dynamic elastic resources.
With the technological revolution brought about by cloud computing, the cloud-native virtual data warehouse PieCloudDB Database enables a transformation from purchasing to leasing in IT systems, delivering on the promises of big data that were unattainable in the PC era. For example, in a bursty scenario like Double 11 Day in China, the required computing power may increase by hundreds of times compared to regular days. The design of PC structures forces customers to invest in hundreds of machines, which are only utilized for a few days out of the 365 days in a year. In such cases, customers have two options: either give up on data computation during bursty scenarios or make substantial upfront investments, which would result in a lower return on investment and missed arbitrage opportunities. Particularly for resource-intensive scenarios like ChatGPT, balancing the resource demands during sudden increases or decreases in website traffic is crucial to effectively utilize resources and control overall expenditures.
In PieCloudDB, storage and computation are treated as two independent variables, each capable of elastic scaling in the cloud. Users can transfer massive amounts of data to the cloud, and the storage in the cloud automatically increases accordingly. This scaling process is handled seamlessly by PieCloudDB without user intervention. If users require greater computing power, they can simply spin up more virtual machines or containers, and PieCloudDB will instantly scale up. Once the bursty computation is completed, users can shut down or downsize the compute cluster, thereby saving on computing costs in the cloud. By decoupling computation from storage, resource pooling is achieved. Users can then lease resources from the pool and pay based on their usage. PieCloudDB allows users to focus on usage without the need to worry about maintenance, upgrades, and other tasks.
In such a system, users continuously store all their data in the cloud, enabling true data sharing among existing and future applications. This helps users realize the true vision of big data, as PieCloudDB makes the promises of big data finally come true.
