RTX Spark 本地 AI 入門:配置差異與本地模型能力
RTX Spark: Local AI First Steps ⭐
Paul Thurrott 介紹 RTX Spark 電腦的本地 AI 入門,並比較三部不同配置電腦的預期用途。128 GB 統一記憶體配置可透過本地智能體運行 120 billion 參數模型,支援最多 one million tokens 的上下文;作者認為,這配置適合要求較高的本地工作負載,而低於 128 GB 的配置則面向創作者、開發者及其他傳統工作站用戶。
Microsoft and Nvidia are promoting the RTX Spark platform as a new class of Windows PCs for local AI, creators, developers, and gaming. Which, when you think about it, is very close to how Copilot+ PC was marketed when the platform first arrived two and a half years ago. But that makes sense, too: As I pointed out in Welcome to the Nvidia RTX Spark Era ⭐, RTX Spark is the third generation AI PC platform, and each of these was a major milestone for local AI on PCs, which each--AI PCs, Copilot+ PCs, and now RTX Spark computers--being a superset of their predecessors in turn.
RTX Spark computers are a lot of things, but it's all rooted in powerful GPU-drive local AI experiences. And that's confusing, frankly, as local AI isn't just behind cloud-based ("frontier") AI functionally, it is much harder to get up and running. With a tool like Copilot, Anthropic Claude, or OpenAI ChatGPT, you just run the app, prompt it with text or your voice, and you're off to the races. With local AI, well, you have a lot of work to do. And while RTX Spark does make a much broader range of local AI models available than is true of other PCs, that also amps up the complexity. It's difficult to know where to start.
But you have to start somewhere. I have three RTX Spark computers in-house for the launch, and they're all a bit different. The Lenovo Pro 9n 15 has the lower-end N1X processor with an 18-core CPU, a 5,120 core GPU, and 32 GB of RAM. The HP OmniBook Ultra 16 has the N1X with the higher-end N1X with a 20-core CPU, 6,144-core CPU and 32 GB of RAM. And then there's the ASUS ProArt P16, which has the higher-end N1X processor and the full 128 GB of unified memory that RTX Spark supports.
My views of this platform are still evolving, of course, but for now, I see this 128 GB configuration as being uniquely suited for the most demanding local workloads--the audience being a new breed of AI enthusiasts and professionals who want the most powerful local AI capabilities in Windows 11--while those NTX Spark configurations with less than 128 GB of RAM are aimed more at creators, developers, and other traditional workstation users. But here's the best news if you've been following the progress of Windows 11 on Arm. Any of these configurations should be solid for gaming, with some obvious exceptions before developers can catch up.
Somewhat more specifically--though, to be fair, this won't mean much to most people--this high-end N1X configuration with 128 GB of unified memory can run 120 billion parameter local models with up to one million tokens of context using local agents. Depending on the use case, this is dramatically more powerful than the smaller local models one will typically run on lower-end RTX Spark configurations, like the other two PCs I'm reviewing, or on other Copilot+ PCs. But this is a complicated enough topic that I'll tackle it over what will likely be at least several articles over the next month or so. Among other concerns, larger...
來源:Thurrott · thurrott.com