Microsoft releases AI chip to compete with Nvidia

Unveiled at the Ignite developer conference in Seattle on November 15, the Azure Maia 100 was introduced by Microsoft as "one of the most advanced chips on the 5-nanometer process" with 105 billion transistors.

According to Rani Borkar, Microsoft's corporate vice president, the company is testing how the Maia 100 can train AI chatbots on the search engine Copilot - the new name for Bing Chat, the GitHub coding assistant Copilot, and also GPT-3.5-Turbo, a large language model from OpenAI.

Microsoft and OpenAI are now trying to attract businesses to use generative AI models. Microsoft CEO Satya Nadella said last month that the field is growing strongly. GitHub Copilot alone grew 40% in the third quarter of 2023 compared to the previous quarter.

“We have more than a million paying Copilot users across more than 37,000 organizations and businesses. That number is growing significantly, especially outside the United States,” Nadella said.

At the event, Microsoft also announced Cobalt 100, aimed at systems that perform common computing tasks and can compete with Intel processors.

The new chip is based on ARM architecture, 64-bit, contains 128 computing cores, high performance operation and the ability to reduce power consumption by 40% compared to other ARM chips that Azure cloud systems are using.

Cobalt is now behind cloud software systems including Microsoft Teams and Azure SQL.

Microsoft is the last of the Big Three cloud providers, along with Google and Amazon, to make its own processors. Google led the way in 2016 with its Tensor Processing Unit (TPU). Amazon followed suit with its Graviton, Trainium, and Inferentia chips.

Despite the new chips, Microsoft said it will continue to work with Nvidia and AMD on Azure chips, with the company planning to add Nvidia’s latest H200 GPU and AMD’s MI300 GPU to its systems next year.

Earlier this week, Nvidia unveiled the H200 chip, the world’s most powerful AI chip. According to the company, testing with Meta’s large Llama 2 language model with 70 billion parameters, the H200 delivered nearly double the performance of the previous most powerful chip, the H100.

Amazon Web Services, Google Cloud, Microsoft Azure and Oracle are the first cloud computing platforms to feature the H200 when it hits the market next year.

In addition to the new chip, Microsoft also announced a custom liquid cooling system for the chip called Sidekicks. In addition to reducing temperatures, this bracket also optimizes the space where the GPU is located.