Bilibili World 2026的英伟达展台，一个"神秘造型"的设备成了全场最火爆的打卡点——RTX Spark真机。
这不是传统意义上的显卡。准确地说，它是一颗把CPU和GPU彻底焊在一起的"AI超级芯片"。具体规格是这样的：20个CPU核心、6144个GPU核心、128GB统一内存、片内带宽达到600GB/s。
128GB统一内存是什么概念？目前消费级显卡的显存天花板是24GB（RTX 4090），就算是专业级的RTX 6000 Ada也只有48GB。128GB意味着什么？意味着你可以在本地跑一个120B参数的大模型，而且是FP8精度，不需要量化、不需要offload、不需要云端API。你的电脑，就是一台小型AI超级计算机。
统一内存：为什么这才是关键突破
很多人看到128GB第一反应是"显存好大"，但真正的革命不是"大"，而是"统一"。
现在的PC架构里，CPU有自己的内存（DDR5），GPU有自己的显存（GDDR6X/HBM）。两边各管各的，数据要在内存和显存之间搬来搬去。你跑AI推理的时候，模型权重先从硬盘加载到系统内存，再从系统内存搬到显存——这个搬运过程就是瓶颈。
RTX Spark的做法是：CPU和GPU共享同一个128GB内存池。通过英伟达自研的第二代互联技术，GPU可以直接访问大容量内存，不需要再搬来搬去。600GB/s的片内带宽是什么水平？目前DDR5-6400的带宽大约是51.2GB/s，HBM3的带宽大约是1.2TB/s左右。600GB/s正好介于两者之间——比传统内存快一个数量级，接近高端显存的水平。
这意味着什么？意味着"端侧大模型"终于不是"只能跑7B小模型"的玩具了。120B参数是什么概念？是GPT-4级别的模型，是Claude Opus级别的模型，是目前最强大的商用模型的参数量级。以前你想跑这种级别的模型，需要一整台8卡H100服务器，现在你只需要一台装了RTX Spark的个人电脑。
个人AI电脑时代，真的要来了
黄仁勋在GTC台北2026上喊出的口号是"新航向、新起点"——重新定义个人电脑。当时很多人觉得这是营销话术，现在真机亮相，大家才意识到：这可能不是口号，是真的要变天了。
过去两年，"端侧AI"一直是个有点尴尬的概念。手机上的端侧AI只能跑个几B的小模型，做个拍照优化、语音助手什么的；PC上的AI加速也就是个NPU，聊胜于无。大家嘴上都说"端侧AI是未来"，但心里都清楚：真正的大模型还得在云端跑。
RTX Spark把这个格局彻底打翻了。如果个人电脑就能跑120B大模型，那云端推理的必要性在哪里？数据隐私问题怎么解？API调用成本还有没有意义？
当然，现在说"端侧取代云端"还为时过早。RTX Spark的价格肯定不会便宜，128GB统一内存的功耗和发热也是问题，模型优化和软件生态更是长路漫漫。但方向已经很清楚了：AI算力正在从云端向端侧扩散，而且扩散的速度比所有人预想的都快。
从数据中心的H100/H200，到企业级的推理芯片，再到消费级的RTX Spark——英伟达正在把AI算力一层一层往下渗透。当120B大模型能在你家电脑上跑的时候，AI产业的整个价值链都会被重写。
明天见。
At NVIDIA's booth at Bilibili World 2026, a "mystery-shaped" device became the hottest photo spot — the real RTX Spark.
This isn't a graphics card in the traditional sense. To be precise, it's an "AI super chip" that fuses CPU and GPU together completely. The specs: 20 CPU cores, 6144 GPU cores, 128GB unified memory, and on-chip bandwidth reaching 600GB/s.
What does 128GB unified memory mean? Today's consumer GPUs top out at 24GB (RTX 4090); even the professional RTX 6000 Ada only has 48GB. 128GB means you can run a 120B-parameter LLM locally — at FP8 precision, no quantization, no offloading, no cloud API needed. Your PC is a small AI supercomputer.
Unified Memory: Why This Is the Real Breakthrough
Many people see 128GB and think "wow, that's a lot of VRAM" — but the real revolution isn't the size, it's the "unified" part.
In today's PC architecture, the CPU has its own memory (DDR5) and the GPU has its own VRAM (GDDR6X/HBM). They manage separately, and data has to be shuttled back and forth between system memory and video memory. When you run AI inference, model weights load from storage to system RAM, then get copied over to VRAM — that transfer process is the bottleneck.
RTX Spark's approach: CPU and GPU share the same 128GB memory pool. Through NVIDIA's proprietary second-generation interconnect technology, the GPU can directly access the large memory pool without all that shuffling. How fast is 600GB/s of on-chip bandwidth? Current DDR5-6400 runs at about 51.2GB/s; HBM3 hits around 1.2TB/s. 600GB/s sits right between them — an order of magnitude faster than traditional memory, approaching high-end VRAM levels.
What does this mean? It means "on-device LLMs" are finally no longer toys limited to 7B small models. 120 billion parameters — that's GPT-4 level, that's Claude Opus level, that's the parameter scale of today's most powerful commercial models. Previously, running models of this caliber required an entire 8x H100 server. Now you just need a PC with an RTX Spark.
The Era of the Personal AI Computer Is Really Coming
Jensen Huang's slogan at GTC Taipei 2026 was "new direction, new starting point" — redefining the personal computer. At the time, many dismissed it as marketing fluff. Now that the real hardware is on display, people are starting to realize: this might not be a slogan. This might be a paradigm shift.
For the past two years, "on-device AI" has been a somewhat awkward concept. On-device AI on phones can only run tiny multi-billion-parameter models, doing photo optimization and voice assistants. AI acceleration on PCs is basically an NPU — better than nothing. Everyone says "on-device AI is the future," but deep down everyone knows: real large models still run in the cloud.
RTX Spark completely upends this格局. If a personal PC can run a 120B LLM, what's the point of cloud inference? How do you solve the data privacy question? Does API call cost even matter anymore?
Of course, it's too early to say "on-device replaces cloud." The RTX Spark definitely won't be cheap. Power consumption and thermal issues with 128GB unified memory are real concerns. Model optimization and software ecosystem are long roads ahead. But the direction is clear: AI compute is diffusing from the cloud to end devices — and it's happening faster than anyone expected.
From data center H100/H200s to enterprise inference chips to consumer-grade RTX Spark — NVIDIA is pushing AI compute downward, layer by layer. When 120B LLMs can run on your home computer, the entire value chain of the AI industry gets rewritten.
See you tomorrow.
当120B大模型能在你家电脑上跑的时候，AI产业的整个价值链都会被重写。
—— Dawn Vision编辑部
When 120B LLMs can run on your home computer, the entire value chain of the AI industry gets rewritten.
— The Dawn Vision Editorial Desk
RTX Spark · NVIDIA · unified memory · 128GB · 120B LLM · on-device AI · personal AI PC · Bilibili World
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎处理的 9 个源信号生成，经编辑部人工审核。素材来源：量子位、36氪、英伟达GTC台北。
This article was generated by the Dawn Vision cognitive engine processing 9 source signals, with human editorial review. Sources: QbitAI, 36Kr, NVIDIA GTC Taipei.
