IT之家 10 月 12 日消息,微软今天正式发布 Maia 300 人工智能芯片。这款机型配备 288GB HBM4 内存,带宽达到 18TB/s,FP4 算力可达 13.8 PFLOPS,基于台积电 N2P 制程工艺构建,计划 2027 年部署。
IT之家注意到,Maia 300 使用多芯粒外观,主要计算部分使用台积电 N2P 2 纳米制程,集成独立内存芯粒和 I/O 芯粒。FP4 算力最高可达 13.8 PFLOPS,FP8 算力为 6.9 PFLOPS。其 SoC 功耗额定值为 950W,显著高于 Maia 200 的 750W。
同时,该机型配备 8 组 HBM4 内存,总容量可达 288GB,带宽为 18TB/s,更合适对内存要求较高的 AI 推理任务。双向扩展互连带宽达到 3.2TB/s,使用基于以太网的双层网络架构。单个机架可容纳 96 颗 Maia 300,最多可组建包含 12000 颗加速器的集群。
本文来源:https://www.ithome.com/1/012/115.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
IT Home, October 12 — Microsoft officially released its Maia 300 AI chip today. The chip features 288GB of HBM4 memory with a bandwidth of 18TB/s, delivers up to 13.8 PFLOPS of FP4 compute performance, and is built on TSMC's N2P process technology, with deployment planned for 2027.
IT Home notes that the Maia 300 features a multi-chip design, with its main compute portion built on TSMC's N2P 2-nanometer process, integrating separate memory and I/O chiplets. It delivers up to 13.8 PFLOPS of FP4 compute and 6.9 PFLOPS of FP8 compute. Its SoC power rating is 950W, significantly higher than the Maia 200's 750W.
Meanwhile, this model is equipped with 8 sets of HBM4 memory, with a total capacity of up to 288GB and a bandwidth of 18TB/s, making it more suitable for AI inference tasks that are memory-intensive. The bidirectional scaling interconnect bandwidth reaches 3.2TB/s, using a dual-layer network architecture based on Ethernet. A single rack can hold 96 Maia 300 chips, and clusters of up to 12,000 accelerators can be assembled.

Source: https://www.ithome.com/1/012/115.htm If you have any copyright concerns, please contact us at 13951431913#139.com or submit a comment on the right.