跳到正文
@kimmonismus· @kimmonismus · X·· 2026-06-02AI 评分48
AI 导读

NVIDIA 在 Computex 发布 RTX Spark ARM 超级芯片:20 核 Grace CPU、6144 CUDA 核心的 Blackwell GPU、最高 128GB 统一内存、1 PetaFLOP FP4,GPU 核心数达 RTX 5070 级别,可本地运行 120B 参数模型。

正文

At Computex, NVIDIA walked into a market it never owned: the PC itself.

RTX Spark is an ARM superchip: 20-core Grace CPU, Blackwell GPU with 6,144 CUDA cores, up to 128GB unified memory, 1 PetaFLOP FP4. The GPU core count sits at RTX 5070 (!) level, in a laptop as thin as a MacBook.

However, the strategy is the story:

NVIDIA used to sell graphics cards. Now it builds the whole system: CPU, GPU, memory and software stack on one package, all on Windows on ARM. That puts it head to head with Intel, AMD, Qualcomm and Apple at the same time. Four rivals, one chip.

The framing is as important as the gaming demos, even with Jensen holding up two laptops running 007 First Light and Forza Horizon 6. NVIDIA calls this a "personal AI computer." The pitch: your PC becomes a teammate, with local agents working next to you while you do something else. 128GB unified memory means 120B-parameter models running locally, no cloud.

NVIDIA is wagering that the PC of the next few years gets built around local AI (!) rather than apps. If that lands, the company has done more than ship a product. It's set the frame everyone else has to compete in.

What's still open: the GPU is power-limited (45–80W), the FPS and battery claims are unbenchmarked, and Windows on ARM still drags a long list of gaming and x86-emulation problems behind it. Ships Fall 2026. Until then, these are slides, not reviews.

This is the most interesting AI launch of the year is a chip.

来源:@kimmonismus · x.com