As companies accelerate the launch of their own chips, including Google’s 7th-generation Tensor Processing Unit (TPU) Ironwood and Amazon’s 3rd-generation Trainium, in addition to Microsoft, the AI chip market that Nvidia had monopolized is entering a full-scale warring states era.

MS announced on Jan. 26 (local time) that it launched the AI chip Maia 200 that enhanced the efficiency of inference operations. According to MS, Maia 200 is a chip manufactured based on TSMC’s 3-nanometer (nm) process, characterized by improving the economics of generating tokens, which are units of AI inference.

MS introduced that Maia 200’s lightweight computing (FP4) performance for AI response generation reaches three times that of Amazon’s proprietary chip Trainium 3rd generation and has higher computational efficiency than Google’s 7th-generation Tensor Processing Unit (TPU) Ironwood.

MS has already installed this chip at its data center in Iowa, United States, and plans to add it to its data center in Arizona so that customers using its cloud service Azure can utilize it in the future.

It is also expected to considerably reduce dependence on Nvidia’s graphics processing units (GPUs).

Accordingly, the AI chip market is shifting from Nvidia’s monopolistic system to an unlimited competitive system among big tech companies such as Google, Meta, and AWS. Even electric vehicle manufacturer Rivian plans to create its own AI chip and apply it to new vehicles scheduled for release this year.

The competitive landscape is expected to become even more intense with OpenAI joining in, having announced its own chip production within the year in partnership with Broadcom.

MS released a software development kit (SDK) along with Maia 200. Reuters analyzed that this SDK was developed to expand the developer ecosystem, targeting Nvidia’s software platform CUDA, where Nvidia maintains dominance.

CUDA is a programming language and tool used by developers to use Nvidia GPUs, and it is evaluated as having a strong lock-in effect to the extent that more than 90% of AI developers worldwide are familiar with CUDA, leading to stories that even if they want to use other chips, they can’t switch because they have to rewrite the code.

However, recently, alternatives aimed at breaking down Nvidia’s monopolistic system, such as JAX, an AI development tool developed by Google, are emerging one after another.

MS Chief Executive Officer (CEO) Satya Nadella emphasized about Maia 200, “This product, designed for the industry’s best inference efficiency, has 30% higher performance per dollar compared to existing systems,” and “it has the best performance among proprietary semiconductors made by hyperscalers (large data center operators).”

Source: BusinessKorea