
structure|1️⃣ 三级笔记、思想框架
DeepSeek 下一代模型 V4 将能在华为最新芯片 Ascend 950PR 上运行——这是中国半导体自主化进程中的一个里程碑事件。围绕这个里程碑,文章展开了三条线索:需求侧的抢购潮、供给侧的技术攻坚、以及地缘政治下的芯片博弈。
concepts|2️⃣ 关键概念、概念网络
context:
The Ascend 950PR delivers roughly 2.8 times the compute performance of Nvidia's H20. The Ascend 950PR was designed specifically for inference.
费曼一下:华为专为 AI 推理设计的最新芯片。算力约为 Nvidia H20 的 2.8 倍,但仍不及 H200。关键设计决策是兼容 Nvidia 编程指令,降低迁移门槛。本月开始量产,是当前中国 AI 公司能大规模获取的最强国产推理芯片。
context:
Rather than competing head-on with Nvidia in model training, Huawei shifted the focus to inference, the process by which AI models generate answers or perform tasks.
费曼一下:AI 模型生成答案或执行任务的过程,与"训练"相对。华为的核心战略选择——不在训练芯片上与 Nvidia 硬碰硬,而是押注推理市场。推理比训练更容易实现、需求量更大,且随着 DeepSeek 等模型的广泛部署,推理算力需求正在爆发式增长。
context:
The new model, called V4, will be able to run on the latest chips designed by Huawei Technologies.
费曼一下:DeepSeek 的下一代 AI 模型,最大意义不在于模型能力本身,而在于它是第一个从设计阶段就考虑华为芯片兼容性的顶级中国大模型。V4 还有两个额外变体,分别针对不同能力优化,全部基于中国芯片运行。
context:
Its chip design arm, HiSilicon, has been developing silicon for more than two decades.
费曼一下:华为的芯片设计部门,二十多年芯片研发积累。产品线包括麒麟手机处理器和昇腾 AI 加速器。是华为在 AI 芯片领域与 Nvidia 竞争的核心技术团队。
context:
Washington also blocked the export of advanced high-bandwidth memory, a critical component in AI processors, to China in 2024. Huawei had already started building domestic HBM chips and has plans to mass-produce them this year.
费曼一下:AI 处理器中的关键零部件,决定芯片能多快地读写数据。2024 年被美国列入对华出口管制。华为已在自研 HBM,但量产尚需时间,早期 Ascend 950PR 仍需使用海外 HBM——这是中国芯片自主化中最后一块难啃的硬骨头之一。
context:
Reacting to Washington's technology embargo, Chinese policymakers sped up efforts that included generous subsidies to homegrown chipmakers, mandating companies to buy local wherever possible, and pressing AI model developers to build models that could run on Chinese silicon.
费曼一下:美国对华技术禁运政策,限制先进芯片制造设备(如 EUV 光刻机)、高端 AI 芯片和 HBM 对华出口。本文的核心背景——正是出口管制催生了中国半导体自主化的加速,也催生了 DeepSeek V4 + 华为芯片这个组合。
context:
Huawei continues to face capacity constraints at the facilities of its partner, Semiconductor Manufacturing International Corp., China's top foundry.
费曼一下:中国最大的芯片代工厂,华为芯片的唯一制造伙伴(因美国制裁无法使用台积电)。SMIC 的产能限制是华为芯片大规模交付的最大瓶颈,也是整个中国 AI 芯片自主化链条中最脆弱的一环。
context:
DeepSeek has spent the past few months working directly with Huawei and another Chinese chip designer, Cambricon Technologies, to complete that work.
费曼一下:中国另一家 AI 芯片设计公司,与华为一起参与了 DeepSeek V4 的芯片适配工作。它的出现说明中国 AI 芯片生态正在形成多元供给格局,而非只有华为一家。
agentic reading|3️⃣ 费曼 x3