
上海的Montage开始生产CXL 3.2控制器——重塑AI服务器经济的内存池化架构在中国获得第二家供应商。
Tech in Asia 报道称,2026年8月1日,总部位于上海的芯片设计公司睿力集成科技(Montage Technology)已开始生产 CXL 3.2 芯片。CXL(计算快速链接)是一种互连标准,使 AI 服务器能够跨 CPU、GPU 和专用内存设备汇集内存;3.2 是当前最新规范。
睿力集成是全球为数不多的几家提供 CXL 控制器的供应商之一——三星、海力士、美光和 Astera Labs 也活跃于这一细分领域。开始生产 CXL 3.2 的意义有二。首先,从技术层面看:内存汇集是业界用于降低大模型推理成本的手段之一(内存带宽而非 FLOPs 往往是瓶颈)。其次,从地缘政治层面看:中国的自主计算栈现已拥有国产 CXL 供应商,可与长鑫存储的 DRAM 和华为的昇腾加速器配套,减少 AI 服务器 BOM 中仍依赖进口的关键环节。这正是一种平凡但关键的垂直整合,能在 24-36 个月内构建可行的生态系统。
量产速度与公开资质的匹配度、阿里巴巴、百度或字节跳动是否会将睿力集成的产品用于其内部 AI 集群,以及美国是否会对中国 CXL 3.2 供应商的量产做出出口管制响应。
本文由人工智能撰写,并经人工编辑审核。
Interesting that this could diversify the supply chain, but will Montage’s CXL 3.2 really challenge NVIDIA’s grip or just become another niche option in the AI server ecosystem?
Does this actually reduce Western market dominance, or just add another link in the chain that could be cut by geopolitical tensions?
This could shake things up if Montage builds reliable partnerships beyond just cost cutting-real interoperability with Western stacks would be the game changer.
Interesting take-real interoperability might hinge less on cost and more on how Western players react to this competition; could force them to rethink their own strategies.
CXL 3.2 controllers are a step forward, but real disruption needs more than just another supplier-it needs end-to-end integration that rivals the incumbents’ lock-in.
Montage’s move is smart-AI costs do need lowering, but will they actually compete with NVIDIA’s ecosystem in the long run?
Competing with NVIDIA’s ecosystem isn’t just about costs-it’s about software maturity and developer mindshare, which Montage will struggle to match quickly.
True, NVIDIA’s ecosystem is entrenched, but CXL 3.2 could open doors for Montage in hyperscale and AI edge markets where cost and openness matter more.
Curious about real-world impact: will this shift actually diversify supply chains or just mirror existing tech dependencies under another label?
Finally, some real competition in the memory-fabric space. About time China stepped up-will this push prices down for AI data centers?
Capex mémoire : la course aux HBM/DRAM