Microsoft has been developing its own AI chip 'Athena' in-house since 2019 and is already testing it with OpenAI's large-scale language models.

The Information reports that Microsoft has been working on developing an AI chip that can be used to train large-scale language models (LLMs) like ChatGPT and GPT-4 since around 2019. According to the report, some employees at Microsoft and OpenAI have already been testing the performance of LLMs using the AI chip under development.

Microsoft Readies AI Chip as Machine Learning Costs Surge — The Information
https://www.theinformation.com/articles/microsoft-readies-ai-chip-as-machine-learning-costs-surge

Microsoft Building Its Own AI Chip on TSMC's 5nm Process | Tom's Hardware
https://www.tomshardware.com/news/microsoft-athena-ai-chip-tsmc

Microsoft reportedly working on its own AI chips that may rival Nvidia's - The Verge
https://www.theverge.com/2023/4/18/23687912/microsoft-athena-ai-chips-nvidia

New Microsoft AI chip no threat to Nvidia, but growing LLM needs drive custom silicon | VentureBeat
https://venturebeat.com/ai/new-microsoft-ai-chip-no-threat-to-nvidia-but-growing-llm-needs-drive-custom-silicon/

Microsoft reportedly tests chips in stealth mode to boost AI performance | Windows Central
https://www.windowscentral.com/software-apps/microsoft-reportedly-tests-chips-in-stealth-mode-to-boost-ai-performance

Microsoft developing its own AI chip - The Information | Reuters
https://www.reuters.com/technology/microsoft-developing-its-own-ai-chip-information-2023-04-18/

At the time of writing, NVIDIA has become a major supplier of AI chips, and companies are flocking to buy NVIDIA GPUs. OpenAI, the developer of the chat AI ChatGPT, has reportedly needed more than 30,000 NVIDIA A100 GPUs to commercialize ChatGPT. The Verge points out that NVIDIA's H100 AI GPU is selling for over $40,000 on the internet auction site eBay, demonstrating the growing demand for high-end AI chips.

The Information reports that Microsoft is shifting to developing its own AI chips rather than using NVIDIA AI chips. According to the report, Microsoft hopes to save on AI chip investments by developing its own AI chips. The project is codenamed 'Athena,' and the latest version is being considered for manufacturing on TSMC's 5nm process.

The Athena chip is reportedly designed to train AI models using LLM and similar software, including making inferences from the data they train on. It's unclear whether Microsoft will offer the chip to customers of its Azure cloud service, but Athena could become widely available within Microsoft and OpenAI as early as 2024.

The reason why OpenAI is included is because Microsoft entered into a long-term partnership with OpenAI in January 2023. It should be noted that Athena has already begun partial operation on the latest LLMs such as GPT-4.

Microsoft announces long-term partnership with OpenAI, investing hundreds of billions of yen - GIGAZINE

Microsoft is expected to benefit more than other companies from the development of AI chips because it has deployed AI-powered features in Bing , Office apps, and GitHub .

Microsoft is not the only company developing its own chips for AI; other major technology companies, including Google, Amazon, and Meta, are also developing their own chips.

However, Microsoft does not seem to view Athena as a complete replacement for NVIDIA GPUs. In March 2023, Microsoft partnered with NVIDIA on ' DGX Cloud .'

Microsoft has been working on ARM-based chips for several years, and in 2020 Bloomberg reported that Microsoft was considering designing ARM-based processors for servers . While these ARM-based chips have yet to appear, Microsoft is working with AMD and Qualcomm to develop custom chips for the Surface Laptop and Surface Pro X.

The Information notes that Athena may remain a sensitive topic, and it's unclear whether the project will develop or remain unannounced. However, Tom's Hardware points out that given the recent developments in generative AI and the popularity of Bing's chat AI, it wouldn't be surprising if Athena were to develop further.