(Photo = Shutterstock)
Microsoft (MS) and Open AI are said to be investing $100 billion (about 134.75 trillion won) to build a supercomputer. This is due to the judgment that the development of artificial general intelligence (AGI) requires a large-scale computing infrastructure 100 times the current size.
The Information, citing three sources, said on the 29th (local time), that Microsoft and Open AI are running a project to build a data center including a supercomputer with millions of AI chips by 2030, also known as ‘Stargate’. it was reported that.
Sources who confirmed Microsoft’s initial cost estimates said the project could cost up to $100 billion. Most of the costs will be covered by MS. About half of the total cost is expected to be used to purchase the Nvidia GPU.
The computer cluster with millions of chips is huge, 100 times larger than the largest existing data center. Meta recently introduced two of the world’s largest data centers for developing ‘Rama 3’ and next-generation models, with each cluster containing 24,576 GPUs.
Additionally, $100 billion is more than three times the amount Microsoft spent on servers, buildings and equipment last year. Microsoft has already announced that investment in this area will increase rapidly this year due to investments in cloud infrastructure and AI.
The two companies began a five-phase project to build infrastructure starting last summer. The first and second step is that MS will provide servers with hundreds of thousands of GPUs to OpenAI this year and next year.
Phase 3 is said to be underway at the moment. In the fourth phase, we plan to provide a supercomputer for open AI that is smaller than the final goal by 2026.
The purpose of this project is to develop OpenAI artificial general intelligence (AGI). Since last year, CEO Altman has continued to argue that having large-scale computing capability is the key to the development of AGI.
Based on a machine learning concept called ‘trial-time computation’, which emerged in November 2022, Altman and OpenAI claim that when you provide AI models with more computing power, models AI is greatly improved and emphasized. Open AI announced this as the ‘law of expansion’ of conversational AI through a research paper.
That is why Altman has been pursuing a global chip network project since early this year and is trying to establish new chip companies with Microsoft and the United Arab Emirates. This is due to the ruling that the necessary infrastructure cannot be secured by supplying Nvidia GPUs alone.
Satya Nadella, CEO of Microsoft (right) and Sam Altman, CEO of OpenAI (Photo = OpenAI)
However, it is known that there are many technical challenges to overcome when building a supercomputer. For this reason, the project was divided into five phases.
First of all, it is noted that the power required to build a supercomputer is currently about 5 gigawatts, which is enough to operate several large-scale data centers, making it difficult to secure power. For this reason, it has been reported that the use of small nuclear power stations has been considered. Both companies have already invested in this area.
In addition, current designs require many AI chips to be installed in a single rack to increase efficiency and speed, which generates tremendous heat. In addition, the network cable that connects millions of GPUs is also considered a key element.
It has not been decided where the supercomputer will be located and whether it will be split into one data center or multiple adjacent data centers. GPU clusters are generally most efficient when placed in the same data center.
Regarding this, Chris Sharp, CTO of data center operator Digital Reality, said, “Projects like this are a critical step in building AGI,” adding, “It may seem unimaginable now, but it won’t seem so great when the supercomputer. actually completed.” “he said.
Microsoft spokesman Frank Shaw would not comment on supercomputing plans, but said, “We are always planning the next generation of infrastructure innovations needed to push the boundaries of AI capabilities.” OpenAI did not issue an official comment.
Meanwhile, according to Bloomberg, Amazon is also expected to invest $150 billion in several data centers over the next 15 years in response to the rapid increase in demand for AI.
Correspondent Lim Da-jun ydj@aitimes.com
#MSOpenAI #building #supercomputer #times #larger #develop #AGI
