Microsoft Announces Maia 300 Inference Accelerator, Targets 2027 Production and Deployment
Редакция ЦОД:Главное
Microsoft has introduced Azure Maia 300, the third generation of its in-house AI accelerator line, with volume production and deployment planned to begin in 2027. The company intends to scale the Maia family to deployments measured in multiple gigawatts for inference across its own products and services.
- Maia 300 is Microsoft’s third-generation in-house Azure Maia accelerator and is designed for AI inference.
- Volume production and deployment are planned from 2027, not yet under way for Maia 300.
- Microsoft targets multiple-gigawatt Maia-family deployments, but has not disclosed sites, timing by location or CAPEX.
- A stated Maia 300 rack configuration combines 96 chips at 167 kW.
- Microsoft’s performance-per-MW and rack-throughput comparisons are vendor claims without a disclosed independent methodology.
Microsoft announced Azure Maia 300 on 12 October as the third generation of its Maia accelerator programme. The accelerator is designed for inference, with volume production and deployment expected to start in 2027.
Microsoft says it intends to scale the Maia family to deployments at a multiple-gigawatt level and position it as a primary inference platform for its own products and services. This is a future deployment ambition rather than commissioned capacity; the company has not disclosed the country, site, schedule or CAPEX allocation for such deployments.
At rack level, Microsoft states that a Maia 300 configuration integrates 96 chips within 167 kW of power consumption. The accelerator uses a multi-chiplet architecture, with its compute chiplet manufactured on TSMC’s N2P process. Maia 300 includes eight HBM4 stacks, with a stated peak memory bandwidth of 18 TB/s. Microsoft cites 27.6 TB of HBM capacity and 1.7 PB/s of aggregate HBM bandwidth for the rack configuration.
Microsoft claims Maia 300 will deliver 40% more tokens per MW than leading GPU systems, and says rack-level token throughput could be up to seven times that of Maia 200. These are company expectations and claims; the available source does not provide an independent comparison methodology.
The company is already deploying Maia 200 in the United States, Europe and Australia. If Microsoft follows through on multi-gigawatt scaling, its proprietary accelerators could become a material factor in the procurement mix for AI infrastructure and in the power demand of its data-center estate. The stated rack configuration also underscores the role of high-density power systems, cooling, memory and Ethernet networking in inference-oriented AI deployments.
The plan points to a major cloud operator targeting multi-gigawatt scale for dedicated AI inference infrastructure. If executed, the architecture’s high rack power density and Microsoft’s performance-per-watt focus could increase the importance of power delivery, cooling, memory and network design in AI data centers.
- TechNews 科技新報 / КитайскийПервоисточник →
Материал подготовлен редакцией ЦОД: Главное на основе указанных первоисточников.