Microsoft has announced that it has built a new supercomputer to power ChatGPT. After investing $1 billion in OpenAI in 2019, the company wanted to build a cutting-edge machine for the AI research startup. However, as reported by the Bloomberg websitethe company didn’t have something that worked for OpenAI, and it wasn’t sure it could build something that big using its Azure cloud service.
Microsoft builds supercomputer to power ChatGPT
OpenAI is increasingly trying to train artificial intelligence programs (a process called machine learning) to create language models. These software consume large volumes of data and develop more and more parameters with the variables that the AI discovered through training and retraining. That is, a large cloud processing power is needed that is capable of providing high performance for a long period.
To be able to meet OpenAI’s needs, Microsoft had to study ways to use several NVIDIA A100 graphics chips and change the way the company positions servers in racks to avoid a power outage. While Microsoft’s executive vice president, responsible for overseeing the cloud and AI, said there was no specific cost for this project, he said it was “probably higher” than several hundred million dollars. Microsoft Azure AI Infrastructure General Manager Nidhi Chappell says:
We built a system architecture that can operate and be reliable at a very large scale. This is what made ChatGPT possible. This is a model that came out of that. There will be many, many more.
Public reaction and the need for a supercomputer
Microsoft’s supercomputer project for OpenAI enabled the release of ChatGPT, which was a success after being made available to the general public. It was so successful that OpenAI’s language model inspired the businesses of other companies such as Google, which launched Bard, and Baidu, which announced ERNIE Bot.
Microsoft currently uses the same computational resources that were created for OpenAI to train its language models with artificial intelligence, including the new Bing search bot. The company is also selling the system to other customers, as Oficina da Net reported. Now, Microsoft is working on developing the next generation AI-powered supercomputer, which is part of its expanded deal with OpenAI, where the company has added an additional $10 billion to its investment. Scott Guthrie says:
We didn’t build something custom for them – it started out as something custom, but we’ve always built it in a way that generalizes it so that anyone who wants to train a large language model can take advantage of the same improvements. It really helped us become a better cloud for AI in general.
The new supercomputer for ChatGPT
Intense training of an AI language model requires many graphical processing units integrated in one place, i.e. we are talking about a supercomputer. However, after ChatGPT was developed, adjustments were needed in the supercomputer developed by Microsoft. The Windows developer uses graphics chips for inference, but these processors are geographically dispersed in more than 60 regions, where the company’s data centers are located.
Microsoft is now adding NVIDIA’s latest graphics chip for AI workloads, the H100, to the AI supercomputer, as well as the latest version of NVIDIA’s Infiniband networking technology, to enable faster data sharing.
