Microsoft introduced Maia 200, a second-generation Artificial Intelligence (AI) chip that is claimed to be the highest performing first-party silicon today. This chip is manufactured in 3nm.
Maia 200 offers a 216GB HBM3e memory system with a transfer speed of up to 7 TB/s. This hardware is also equipped with an FP8/FP4 native tensor core that is capable of keeping large AI models working optimally and quickly.
The performance of this chip answers the challenges of its competitors in the global market, including Vera Rubin from Nvidia. The FP4 capability of Maia 200 is three times stronger than Amazon Trainium 3, while its FP8 performance is above Google's seventh-generation TPU.
"Maia 200 is also the most efficient inference system Microsoft has ever implemented, with 30% better performance per dollar than the latest generation of hardware in our current fleet," Microsoft explained in its official blog, quoted on Tuesday, January 27.
Each chip can hold more than 140 billion transistors to handle complex low-precision computing. Maia 200 is capable of delivering more than 10 petaFLOPS performance at 4-bit precision to provide more space for more massive AI models.
Maia 200 has a second-generation closed-loop liquid cooling system that ensures the device temperature is maintained. Thanks to the mature end-to-end system validation, the chip deployment time from the factory to the data center can be twice as fast.
Microsoft explained that this chip will support various advanced AI models, one of which is GPT-5.2 from OpenAI. Microsoft's internal team will also use it for the creation of synthetic data and the learning process of strengthening the next generation of models.
This chip has already been used in a data center called US Central which is located in Iowa. Microsoft will also use this chip in the US West 3 region which is in Arizona. It is not yet known for sure when the second location will use this chip.
The English, Chinese, Japanese, Arabic, and French versions are automatically generated by the AI. So there may still be inaccuracies in translating, please always see Indonesian as our main language. (system supported by DigitalSiber.id)
