微软推出第三代Azure Maia 300芯片架构，采用软件定义数据流架构（SDLA）和多芯粒设计，专为云规模的智能体推理工作负载进行优化1。相比上一代Maia 200，该芯片在多项性能指标上实现显著提升：机架级令牌吞吐量最高提升7倍，性能功耗比提升超过2.5倍1。
新一代架构在密度和带宽方面也取得进展1。单个机架集成96个加速器，相比Maia 200加速器密度实现翻倍1。单个加速器的HBM带宽提升2.5倍，容量增加33%，同时计算FLOPs提升36%，Scale-up带宽提升14%1。该芯片采用TSMC N2P工艺制造1，内存架构由UMA改为NUMA跨芯粒设计1。
在部署规模上，Maia 300支持构建包含最多12,288个加速器的超级集群1。单个机架的功耗为167kW，其中超过80%的热负荷通过液冷散热处理1。微软计划在2026年Microsoft Ignite大会上分享更多AI战略信息1。
Microsoft has released the third-generation Azure Maia 300 chip architecture, engineered for cloud-scale intelligent agent inference workloads with a software-defined dataflow architecture (SDLA) and multi-chiplet design.1
The new generation achieves significant performance improvements over its predecessor. Maia 300 delivers up to 7 times greater token throughput at the rack level compared to Maia 200, while improving performance-per-watt efficiency by more than 2.5 times.1 A single rack integrates 96 accelerators—double the density of previous generations—with each accelerator featuring 2.5 times greater HBM bandwidth and 33 percent increased capacity.1 Computational performance increases by 36 percent, and scale-up bandwidth improves by 14 percent.1
The architecture employs TSMC's N2P process technology for compute chips and transitions from a Uniform Memory Access (UMA) to a Non-Uniform Memory Access (NUMA) cross-chiplet architecture.1 Individual racks consume 167 kilowatts of power, with liquid cooling handling more than 80 percent of the thermal load.1 The system supports supercluster configurations with up to 12,288 accelerators.1 Microsoft plans to share additional AI strategy information at the 2026 Microsoft Ignite conference.1
评论
还没有评论，欢迎留下第一条。