GPU
Now Reading
Microsoft Is Building AI Chip Of Its Own To Reduce Reliance On NVIDIA
Contents
0

Microsoft Is Building AI Chip Of Its Own To Reduce Reliance On NVIDIA

by Low Boon ShenApril 19, 2023
What's your reaction?
Me Gusta
0%
WOW
0%
Potato
0%
Sad Reacc
0%
Angery
0%

Microsoft Is Building AI Chip Of Its Own To Reduce Reliance On NVIDIA

While the AI boom has brought NVIDIA huge customer orders, Microsoft is clearly avoiding putting too many eggs in the wrong basket.

Microsoft Is Building AI Chip Of Its Own To Reduce Reliance On NVIDIA

The Information reports Microsoft is designing its own AI chip for its own use, reducing reliance on NVIDIA for applications such as ChatGPT, which requires huge amounts of GPU to run (in the tens of thousands, usually). However, this technically isn’t new – the report claims Microsoft has been secretly developing its AI chips all the way back in 2019 under the codename “Athena”, and some of them are currently being tested by OpenAI as well.

Currently, major AI providers – including OpenAI – gets their compute resource from NVIDIA’s A100 GPUs and the new H100 GPU (based on datacenter-specific Hopper architecture), which is apparently listed as $40,000 on eBay, though it’s hard to determine exact pricing as B2B bulk orders may differ on a case-by-case basis. The recent trend of major tech companies bringing chip design in-house (such as Meta, Google and Amazon Graviton) has certainly give Microsoft a good reason to do the same.

Microsoft certainly has experience in designing custom chips, however. The Microsoft SQ1, based on Qualcomm’s Snapdragon chips, are among the first ARM laptops in to ever released to the market (though it was grossly underpowered as a first-generation product). This time around though, Athena is set to cut overall costs in buying, and possibly maintaining GPGPUs such as NVIDIA A100 – though it is said that the new chips won’t be a direct replacement to NVIDIA’s parts.

Source: The Verge | Tom’s Hardware

Pokdepinion: Not surprising considering a dedicated accelerators is better than a GPGPU for a specialized purpose.

About The Author
Low Boon Shen
Is technology powered by a series of tubes?