NVIDIA 的 DGX Station 有望实现本地万亿参数推理。 The memory math is the catch.
10 月 7 日展示的 NVIDIA 适用于 Windows 的 DGX Station 配备了 748GB 相干内存和高达 20 petaflops，旨在为接近万亿个参数的模型提供本地推理。 It's due in the fourth quarter.
陷阱是内存分裂。报告显示，一旦 Windows 占据份额，128GB 统一内存 SKU 只会向 GPU 暴露约 110GB 的空间。一旦权重、缓存和运行时竞争其余部分，这将决定哪个模型真正适合。
因此，请阅读您真正想要购买的规格，而不是标题。万亿参数的说法是针对胖SKU的；你的桌子可能会变小。
老实说，对于无法将数据发送到云端的团队来说，它是一个真正的工具。作为一种身份象征，这是一种运行适合更便宜装备的模型的昂贵方式。
Every rank on Best AI Utility is paid and time-stamped — outbid the current holder with one crypto payment and the rank is yours until someone outbids you.
人工智能工具太多？这是找到您真正需要的东西的更好方法
If you only check AI news on weekends, this was the weekend to check. The theme was not a bigger model, it was models that do a specific job well enough to hand them the keys.
A model that finds vulnerabilities but refuses to step outside authorized scope is not just a better scanner. It is the first time the powerful exploit-finder has been gated to the people who already had permission to look.
