英伟达 GTC 大会已经成了 AI 界超级碗，没有剧本也没有提词器，中途黄仁勋被线缆卡住，反而是这场高浓度 AI 发布会里最有人味的片段，在当今提前基本提前彩排或录播的科技发布会里已经很稀缺了。

刚刚，黄仁勋再次发布了全新一代核弹级 AI 芯片，不过这场发布会的还有个隐藏主角——DeepSeek。

由于智能体 AI（Agentic AI）和推理能力的提升，现在所需的计算量至少是去年此时预估的 100 倍。

推理成本效率给 AI 行业带来影响，而不是简单地堆积计算能力，成为贯穿这场发布会的主线。英伟达要变成 AI 工厂，让 AI 以超越人类的速度学习和推理。

推理本质上是一座工厂在生产 token，而工厂的价值取决于能否创造收入和利润。因此，这座工厂必须以极致的效率打造。

黄仁勋掏出的英伟达新「核弹」也在告诉我们，未来的人工智能竞争不在于谁的模型更大，而在于谁的模型具有最低的推理成本和更高推理的效率。

除了全新 Blackwell 芯片，还有两款「真·AI PC」

全新的 Blackwell 芯片代号为「Ultra」，也就是 GB300 AI 芯片，接棒去年的「全球最强 AI 芯片」B200，再一次实现性能上的突破.

Blackwell Ultra 将包括英伟达 GB300 NVL72 机架级解决方案，以及英伟达 HGX B300 NVL16 系统。

Blackwell Ultra GB300 NVL72 将于今年下半年发布，参数细节如下：

• 1.1 EF FP4 Inference：在进行 FP4 精度的推理任务时，能够达到 1.1 ExaFLOPS（每秒百亿亿次浮点运算）。
• 0.36 EF FP8 Training：在进行 FP8 精度的训练任务时，性能为 1.2 ExaFLOPS。
• 1.5X GB300 NVL72：与 GB200 NVL72 相比，性能为 1.5 倍。
• 20 TB HBM3：配备了 20TB HBM 内存，是前代的 1.5 倍
• 40 TB Fast Memory：拥有 40TB 的快速内存，是前代的 1.5 倍。
• 14.4 TB/s CX8：支持 CX8，带宽为 14.4 TB/s，是前代的 2 倍。

单个 Blackwell Ultra 芯片将和前代一样提供相同的 20 petaflops（每秒千万亿次浮点运算） AI 性能，但配备更多的 288GB 的 HBM3e 内存。

如果说 H100 更适合大规模模型训练，B200 在推理任务中表现出色，那么 B300 则是一个多功能平台，预训练、后训练和 AI 推理都不在话下。

英伟达还特别指出，Blackwell Ultra 也适用于 AI 智能体，以及用于训练机器人和汽车自动驾驶的「物理 AI」。

为了进一步增强系统性能，Blackwell Ultra 还将与英伟达的 Spectrum-X 以太网和英伟达 Quantum-X800 InfiniBand 平台集成，为系统中的每个 GPU 提供 800Gb/s 的数量吞吐量，帮助 AI 工厂和云数据中心能够更快处理 AI 推理模型。

除了 NVL72 机架，英伟达还推出了包含单个 GB300 Blackwell Ultra 芯片的台式电脑 DGX Station。Blackwell Ultra 之外，这个主机还将配备 784GB 的同一系统内存，内置 800Gbps 英伟达 ConnectX-8 SuperNIC 网络，能够支持 20 petaflops 的 AI 性能。

而之前在 CES 2025 展示的「迷你主机」Project DIGITS 也正式被命名为 DGX Spark，搭载专为桌面优化的 GB10 Grace Blackwell 超级芯片，每秒可提供高达 1000 万亿次 AI 计算操作，用于最新 AI 推理模型的微调和推理，包括 NVIDIA Cosmos Reason 世界基础模型和 NVIDIA GR00T N1 机器人基础模型。

黄仁勋表示，借助 DGX Station 和 DGX Spark，用户可以在本地运行大模型，或者将其部署在 NVIDIA DGX Cloud 等其他加速云或者数据中心基础设施上。

这是 AI 时代的计算机。

DGX Spark 系统现已开放预订，而 DGX Station 预计将由华硕、戴尔、惠普等合作伙伴于今年晚些时候推出。

下一代 AI 芯片 Rubin 官宣，2026 年下半年推出

英伟达一直以科学家的名字为其架构命名，这种命名方式已成为英伟达文化的一部分。这一次，英伟达延续了这一惯例，将下一代 AI 芯片平台命名为「Vera Rubin」，以纪念美国著名天文学家薇拉·鲁宾（Vera Rubin）。

黄仁勋表示，Rubin 的性能将达到 Hopper 的 900 倍，而 Blackwell 相较 Hopper 已实现了 68 倍的提升。

其中，Vera Rubin NVL144 预计将在 2026 年下半年发布。参数信息省流不看版：

• 3.6 EF FP4 Inference：在进行 FP4 精度的推理任务时，能够达到 3.6 ExaFLOPS（每秒百亿亿次浮点运算）。
• 1.2 EF FP8 Training：在进行 FP8 精度的训练任务时，性能为 1.2 ExaFLOPS。
• 3.3X GB300 NVL72：与 GB300 NVL72 相比，性能提升了 3.3 倍。
• 13 TB/s HBM4：配备了 HBM4，带宽为 13TB/s。
• 75 TB Fast Memory：拥有 75 TB 的快速内存，是前代的 1.6 倍。
• 260 TB/s NVLink6：支持 NVLink 6，带宽为 260 TB/s，是前代的 2 倍。
• 28.8 TB/s CX9：支持 CX9，带宽为 28.8 TB/s，是前代的 2 倍。

标准版 Rubin 将配备 HBM4，性能比当前的 Hopper H100 芯片大幅提升。

Rubin 引入名为 Grace CPU 的继任者——Veru，包含 88 个定制的 Arm 核心，每个核心支持 176 个线程，并通过 NVLink-C2C 实现 1.8 TB/s 的高带宽连接。

英伟达表示，定制的 Vera 设计将比去年 Grace Blackwell 芯片中使用的 CPU 速度提升一倍。

与 Vera CPU 搭配时，Rubin 在推理任务中的算力可达 50 petaflops，是 Blackwell 20 petaflops 的两倍以上。此外，Rubin 还支持高达 288GB 的 HBM4 内存，这也是 AI 开发者关注的核心规格之一。

实际上，Rubin 由两个 GPU 组成，而这一设计理念与当前市场上的 Blackwell GPU 类似——后者也是通过将两个独立芯片组装为一个整体运行。

从 Rubin 开始，英伟达将不再像对待 Blackwell 那样把多 GPU 组件称为单一 GPU，而是更准确地按照实际的 GPU芯 片裸片数量来计数。

互联技术也升级了，Rubin 配备第六代 NVLink，以及支持 1600 Gb/s 的 CX9 网卡，能够加速数据传输并提升连接性。

除了标准版 Rubin，英伟达还计划推出 Rubin Ultra 版本。

Rubin Ultra NVL576 则将于 2027 年下半年推出。参数细节如下：

• 15 EF FP4 Inference：在 FP4 精度下进行推理任务时，性能达到 15 ExaFLOPS。
• 5 EF FP8 Training：在 FP8 精度下进行训练任务时，性能为 5 ExaFLOPS。
• 14X GB300 NVL72：相比 GB300 NVL72，性能提升 14 倍。
• 4.6 PB/s HBM4e：配备 HBM4e 内存，带宽为 4.6 PB/s。
• 365 TB Fast Memory：系统拥有 365 TB 的快速内存，是前代的 8 倍。
• 1.5 PB/s NVLink7：支持 NVLink 7，带宽为 1.5 PB/s，是前代的 12 倍。
• 115.2 TB/s CX9：支持 CX9，带宽为 115.2 TB/s，是前代的 8 倍。

在硬件配置上，Rubin Ultra 的 Veras 系统延续了 88 个定制 Arm 核心的设计，每个核心支持 176 个线程，并通过 NVLink-C2C 提供 1.8 TB/s 的带宽。

而 GPU 方面，Rubin Ultra 集成了 4 个 Reticle-Sized GPU，每颗 GPU 提供 100 petaflops 的 FP4 计算能力，并配备 1TB 的 HBM4e 内存，在性能和内存容量上都达到了新的高度。

为了在瞬息万变的市场竞争中站稳脚跟，英伟达的产品发布节奏已经缩短至一年一更。发布会上，老黄也正式揭晓下一代 AI 芯片的命名——物理学家费曼(Feynman)。

随着 AI 工厂的规模不断扩大，网络基础设施的重要性愈发凸显。

为此，英伟达推出了 Spectrum-X 和 Quantum-X 硅光网络交换机，旨在帮助 AI 工厂实现跨站点连接数百万 GPU，同时显著降低能耗和运营成本。

Spectrum-X Photonics 交换机具有多种配置，包括：

• 128 端口 800Gb/s或 512 端口 200Gb/s 配置，总带宽达 100Tb/s
• 512 端口 800Gb/s或 2048 端口200Gb/s配置，总吞吐量达 400Tb/s

与之配套的 Quantum-X Photonics 交换机则基于 200Gb/s SerDes 技术，提供 144 端口 800Gb/s 的 InfiniBand 连接，并采用液冷设计高效冷却板载硅光子组件

与上一代产品相比，Quantum-X Photonics 交换机为 AI 计算架构提供 2 倍速度和 5 倍可扩展性。

Quantum-X Photonics InfiniBand 交换机预计于今年晚些时候上市，而 Spectrum-X Photonics 以太网交换机预计将于 2026 年推出。

随着 AI 的快速发展，对数据中心的带宽、低延迟和高能效需求也急剧增加。

英伟达 Spectrum-X Photonics 交换机采用了一种名为 CPO 的光电子集成技术。其核心是将光引擎（就是能处理光信号的芯片）和普通的电子芯片（比如交换芯片或 ASIC 芯片）放在同一个封装里。

这种技术的好处很多：

• 传输效率更高：因为距离缩短，信号传输更快。
• 功耗更低：距离短了，传输信号需要的能量也少了。
• 体积更小：把光和电的部件集成在一起，整体体积也变小了，空间利用率更高。

AI 工厂的「操作系统」Dynamo

未来将没有数据中心，只有 AI 工厂。

黄仁勋表示，未来，每个行业、每家公司拥有工厂时，都将有两个工厂：一个是他们实际生产的工厂，另一个是 AI 工厂，而 Dynamo 则是专门为「AI 工厂」打造的操作系统。

Dynamo 是一款分布式推理服务库，为需要 token 但又无法获得足够 token 的问题提供开源解决方案。

简单来说，Dynamo 有四个方面的优势：

•  GPU 规划引擎，动态调度 GPU 资源以适应用户需求
• 智能路由器，减少 GPU 对重复和重叠请求的重新计算，释放更多算力应对新的传入请求
• 低延迟通信库，加速数据传输
• 内存管理器，智能在低成本内存和存储设备中的推理数据

人形机器人的露脸环节，永远不会缺席

人形机器人再一次成为了 GTC 大会的压轴节目，这次英伟达带来了 Isaac GR00T N1，全球首款开源人形机器人功能模型。

黄仁勋表示，通用机器人技术的时代已经到来，借助 Isaac GR00T N1 核心的数据生成以及机器人学习框架，全球各地的机器人开发人员将进入 AI 时代的下一个前沿领域。

这个模型采用「双系统」架构，模仿人类的认知原理：

• 系统 1：快速思考的动作模型，模仿人类的反应或直觉
• 系统 2：慢思考的模型，用于深思熟虑的决策。

在视觉语言模型的支持下，系统 2 对环境和指令进行推理，然后规划动作，系统 1 将这些规划转化为机器人的的动作。

GR00T N1 的基础模型采用广义类人推理和技能进行了预训练，而开发人员可以通过真实或合成数据进行后训练，满足特定的需求：既可以完成工厂的特定任务，也可以在家里自主完成家务。

黄仁勋还宣布了与 Google DeepMind 和 Disney Research 合作开发的开源物理引擎 Newton。

一台搭载 Newton 平台的机器人也登上了舞台，黄仁勋称之为「Blue」，外观神似《星球大战》中的 BDX 机器人，能够用声音和动作和黄仁勋互动。

8 块 GPU，DeepSeek-R1 推理速度创全球之最

英伟达实现了全球最快的 DeepSeek-R1 推理。

官网显示，一台搭载 8 个 Blackwell GPU 的 DGX 系统，在运行 6710 亿参数的 DeepSeek-R1 模型时，可实现每用户每秒超过 250 个 token 的速度，或达到最高吞吐量每秒超过 30000 个 token。

通过硬件和软件的结合，自今年 1 月以来，英伟达在 DeepSeek-R1 671B 模型上的吞吐量提升了约 36 倍，每 token 的成本效率提高了约 32 倍。

为了实现这一成就，英伟达完整的推理生态系统已针对 Blackwell 架构进行了深度优化，不仅整合 TensorRT-LLM、TensorRT Model Optimizer 等先进工具，还无缝支持 PyTorch、JAX 和 TensorFlow 等主流框架。

在 DeepSeek-R1、Llama 3.1 405B 和 Llama 3.3 70B 等模型上，采用 FP4 精度的 DGX B200 平台相较于 DGX H200 平台，推理吞吐量提升超过 3 倍。

值得注意的是，此次发布会的主题演讲并未提及量子计算，但英伟达特意在这届 GTC 大会设置了量子日，邀请了多家当红量子计算公司的 CEO 出席。

要知道黄仁勋年初一句「量子计算还需 20 年才实用」的论断犹在耳畔。

一改口风的背后，离不开微软耗时 17年研发的拓扑量子芯片 Majorana 1 实现 8 个拓扑量子比特集成，离不开 Google Willow 芯片宣称用 5 分钟完成经典计算机需 10^25 年处理的任务，推动了量子计算的热潮。

芯片无疑是重头戏，但一些软件的亮相同样值得关注。

硅谷著名投资人马克·安德森曾提出软件正在吞噬世界（Software is eating the world）的论断，其核心逻辑在于软件通过虚拟化、抽象化和标准化，正在成为控制物理世界的基础设施。

不满足于做「卖铲人」，英伟达的野心是打造 AI 时代的「生产力操作系统」。从汽车智能驾驶，到制造业的数字孪生工厂，这些贯穿整场发布会的案例都是将 GPU 算力转化为行业生产力的具象化表达。

实际上，无论是发布会上亮相的最新核弹芯片，还是押注战未来的量子计算，黄仁勋在这场发布会上对 AI 未来发展的洞察和布局，都比当下的技术参数与性能指标更具看点。

在介绍 Blackwell 与 Hopper 架构的对比时，黄仁勋还不忘幽默一把。

他以一个 100MW 工厂的对比数据为例，指出采用 Hopper 架构需要 45,000 颗芯片和 400 个机架，而 Blackwell 架构凭借更高的效率显著减少了硬件需求。

于是，黄仁勋那句经典的总结再次抛出，「the more you buy, the more you save」（买得越多，省得越多）。」随后话锋一转，他又补充说，「the more you buy, the more you make」（买得越多，赚得越多）。

随着 AI 领域的重心从训练转向推理，英伟达更需要证明其软硬件生态在推理场景的不可替代性。

一方面，Meta、Google 等巨头自研 AI 芯片，可能分流 GPU 市场需求。

另一方面，英伟达最新 AI 芯片的适时亮相，回应如 DeepSeek 的开源模型对 GPU 需求的冲击，并展示推理领域技术优势，也是为了对冲市场对训练需求见顶的担忧。

最近估值跌至 10 年低位的英伟达，比以往任何时候都需要一场酣畅淋漓的胜利。

#欢迎关注爱范儿官方微信公众号：爱范儿（微信号：ifanr），更多精彩内容第一时间为您奉上。

爱范儿 | 原文链接 · 查看评论 · 新浪微博

This is the second of this weekend’s two articles in which the artists and paintings of Ukraine tell their own story. Each of the links given takes you back to the series of articles I compiled here a couple of years ago.

By the end of the nineteenth century, Ukrainian art schools were at last training the Ukrainian artists of the future, who were able to make their own styles and develop distinctive movements. Among them were some who went on to earn a place internationally. Unfortunately, history conspired to change all this in the political unrest around the October 1917 Revolution in Russia, and through two World Wars. As a result, the lives of many Ukrainian artists were brought to an early end, by disease, starvation, or execution. A large proportion of their output has been deliberately suppressed, hidden away in collections of banned works, destroyed, or looted.

Mykola Pymonenko (1862–1912) was born in a village outside Kyiv and started his training in his father’s icon workshop in the city, prior to his discovery by Mykola Murashko of the Kyiv Art School, where he trained before heading off to Saint Petersburg. He returned to Kyiv in 1884 to teach and to paint in the Naturalist style that was so popular at the Salon in Paris at the time. It was he who perhaps painted the first distinctively Ukrainian works that drew on local themes such as Paska at Easter, traditional weddings and the grain harvest. Kazymyr Malevych was among his pupils.

Mykola Pymonenko

Many Ukrainian artists had depicted Zaporozhian Cossacks, and they remained a favourite theme for Ilia Repin right up to his death, but the first prominent specialist national history painter was Mykola Ivasyuk (1865–1937), who was born and brought up in Zastavna in western Ukraine when it was still part of the Austro-Hungarian Empire. He therefore trained at the Academy of Fine Arts in Vienna, but chose to spend much of his career in Chernivtsi and Kyiv, where he also taught.

During the nineteenth century, the evolution of painting in Ukraine had largely been constrained by the orthodoxy of the Imperial Academy, and the supply of Russian patrons. As its own art schools flourished, and support for the arts grew, the pace of progress rose rapidly in Ukraine. The first decade of the twentieth century saw Ukrainian artists in the same avant garde as those in France and the rest of Europe, until the First World War and the October 1917 Revolution.

The war threw Ukraine into the midst of the conflict between Austro-Hungary and Russia, with Ukrainians fighting one another on behalf of two different empires. Then from 1917 onwards, the country lapsed in and out of complete chaos. By 1920, many Ukrainian artists had been forced to leave, or were intending to do so.

For painters like Ivasyuk, the world changed too rapidly. Initially he was commissioned to design postage stamps, then in 1926 he was made a professor at the Kyiv Art Institute. In a few years he had been moved away to Odesa as a result of political criticism. In the autumn of 1937, he was arrested, convicted of terrorism on the basis of his history paintings, and was shot by a firing squad at the age of seventy-two. Many of his paintings were confiscated or destroyed in a bid to erase him and his work completely.

Mykola Ivasyuk

Viktor Zarubin (1866–1928) was born in Kharkiv, trained under Arkhyp Kuindzhi, and painted extensively in Ukraine and northern France.

Viktor Zarubin

Ivan Trush (1869–1941) was born in Vysotsko, to the north-east of Lviv. He settled in the city, where he was involved in the establishment of Lviv National Museum. He was a prominent portraitist, and painted Impressionist landscapes.

Ivan Trush

Petro Nilus (1869–1943) was born in Balta, in south-west Ukraine, and moved to Odesa, where he studied under Kyriak Kostandi. He was active in Odesa for much of his career until moving to Paris in 1920.

Petro Nilus

Mykola Burachek (1871-1942) was born in Letychiv, western Ukraine, and trained under Khariton Platonov in Kyiv. He taught in Kyiv from 1917, then in Kharkiv from 1925. He died there during its Nazi occupation.

Mykola Burachek

Oleksandr Murashko (1875–1919) was born in Kyiv, where he became a major figurative painter. He taught there from 1909, and was a co-founder of the Ukrainian State Academy of Arts. He was shot dead by a street gang.

Oleksandr Murashko

Kazimierz Sichulski (1879–1942) was born in Lviv. He travelled in the Carpathian Mountains from 1905, where he painted Hutsul peoples. He taught in Lviv from 1918, then in Kraków, Poland from 1930, before returning to Lviv in 1939.

Kazimierz Sichulski’s Galician Landscapes 1
Kazimierz Sichulski’s Galician Landscapes 2

Fedir Krychevskyi (1879–1947) was born in Lebedyn, near Sumy in north-east Ukraine. He studied with Gustav Klimt in Vienna before returning to teach in Kyiv in 1914, where he was appointed Rector of the Ukrainian State Academy of Arts. He remained in Kyiv during the Second World War, but was arrested by Soviet forces in 1943, and died of starvation in Irpin during the famine of 1947.

Fedir Krychevskyi

Kazymyr Malevych (1879–1935) was born in Kyiv of Polish descent, where he started his studies. He became a Cubo-Futurist by 1912, then went on to Suprematism. He taught at Kyiv Art Institute from 1928 alongside Oleksandr Bohomazov, but was sacked from there in 1930, was arrested by the KGB and threatened with execution.

Kazymyr Malevych

Oleksandr Bohomazov (1880–1930) was born in Yampil, near Lyman in east Ukraine. He trained under Mykola Pymonenko at the Kyiv Academy of Arts from 1902, alongside Oleksandra Ekster and the sculptor Oleksandr Arkhypenko who were to play major roles in the development of modernist art in Ukraine and Europe. After a period studying in Moscow, he returned to Kyiv in 1908, where he became one of the leaders of the avant garde. In 1914 he wrote an innovative treatise on modern painting that formed the basis of his teaching at the Kyiv Art Institute from 1922.

Oleksandr Bohomazov

Like Ivasyuk, Mykhailo Boichuk (1882-1937) came from Galicia in Austro-Hungary, where he was born to the south of Ternopil, but trained first in Lviv then in the Academy of Fine Arts in Kraków, Poland. He returned to Lviv in 1910, where he developed a novel style, unique to Ukraine, known as Monumentalism or Boichukism, which brings together traditional Byzantine icon painting and the pre-Renaissance. This enjoyed recognition and popularity during the 1920s, when there were more than two dozen visual artists creating commissioned works for public buildings throughout Ukraine. After the October 1917 Revolution he was a co-founder of the Ukrainian State Academy of Arts.

Boichuk and his colleagues fell from grace during Stalin’s Great Purge of 1937, when he was accused of being a bourgeois nationalist. For that, he, his wife and many of his colleagues were executed that year, and most of their work destroyed. They weren’t the last to die for their art, either: in 1946, for instance, Ivan Ivanets, director of the Lviv Art Gallery, was kidnapped and killed in Russia.

Mykhailo Boichuk

Expatriates of this period include Oleksandr Shevchenko, Oleksandra Ekster, Arnold Lakhovskyi, Wladimir Baranoff-Rossiné and Abraham Mintchine.

Oleksandr Shevchenko (1882-1948) was born in Kharkiv, and painted for much of his career in Moscow.

Oleksandr Shevchenko

Oleksandra Ekster (1882–1949) was born in Poland in a Belarusian family but trained under Mykola Pymonenko in Kyiv, where she launched her career. She lived in Paris from 1906, where she developed Cubo-Futurism before returning to Kyiv in 1914, where she opened an art school in 1918. She then went to Odesa before going to Moscow, and migrated to Paris in 1924.

Oleksandra Ekster

Arnold Lakhovskyi (1880–1937) was born in Chornobyl, in north Ukraine. He trained in Odesa and Munich, and moved to Paris in 1925, then to New York City in 1933, where he was a successful portraitist.

Arnold Lakhovskyi

Wladimir Baranoff-Rossiné (1888-1944) was born in Kherson, and studied in Odesa. He moved to Paris in 1910. During the First World War he lived in Nordic countries, then taught in Moscow in 1920. He was a prolific inventor, whose inventions include the optophonic piano, and he was an early developer of military camouflage. He moved back to Paris in 1925, where he was arrested by the Gestapo in 1943, and he died in Auschwitz.

Wladimir Baranoff-Rossiné

Abraham Mintchine (1898-1931) was born in Kyiv, where he studied at its Art School and with Oleksandra Ekster. He left for Berlin in 1923, then lived in Paris from 1925, where he painted prolifically before he died suddenly in 1931.

Abraham Mintchine

Despite attempts at their assimilation, control and waves of destruction under a succession of empires, painting in Ukraine has somehow flourished as a result of the dedication of this succession of artists. Long may the artists, teachers and art collections of Ukraine flourish.

References

Andrey Kurkov and others (2022) Treasures of Ukraine, A Nation’s Cultural Heritage, Thames & Hudson. ISBN 978 0 500 02603 8.
Konstantin Akinsha and others (2022) In the Eye of the Storm, Modernism in Ukraine 1900-1930s, Thames & Hudson. ISBN 978 0 500 29715 5.