And hence it divulged the “Athena” Maia 100 XPU in November 2023 with very few details and some rack photos. Perhaps not, because the Braga Maia 200 chip just does FP4 and FP8 on the tensor units and BF16 and FP32 on the vector units. With the Maia 200, Microsoft is stepping up to TSMC’s N3P performance variant of its 3 nanometer process to etch the chip. Here is an empty pair of Maia 200 racks with a coolant distribution rack on the left:And here are some half populated Maia 200 racks in Microsoft’s Azure datacenters:The US Central region of the Azure cloud, which is located outside of Des Moines, Iowa has Maia 200 racks right now, and the US West 3 region outside of Phoenix, Arizona will get them next. AI researchers at Microsoft will also use the Maia 200 to generate synthetic data for training in-house models.