NVIDIA 发布面向智能体工作负载的端到端硬件栈
NVIDIA 不只卖卡，还按任务配好了模型家族，Nemotron 后训练后能超过前沿模型，小模型路线挺省钱。
NVIDIA 在 Jensen 的主题演讲中展示了面向智能体（agentic）工作负载的端到端硬件栈，覆盖自动驾驶、蛋白质发现等多个任务方向。针对不同任务，NVIDIA 提供了对应的模型家族，硬件设计也随之跟进，部分模型目前在各自领域达到 SOTA 水平。演讲还强调了对显卡实际使用方式的关注，例如为新卡群配备传感器。此外，NVIDIA 对 Nemotron 进行后训练（post-training），使其在多项任务上超过前沿模型，作者认为用小模型完成特定任务比一味使用前沿模型更省成本也更环保。
I loved Jensen's keynote yesterday
@NVIDIAAI has built end to end hardware stack for agentic workloads but also has model families for each task you want to build (hardware also follows them) whether it's self-driving, discovering proteins etc and so many of them are sota atm
they care a ton about how their cards are used (e.g. cards with sensors for new fleets)
he also highlighted post-training Nemotron for various tasks surpassing frontier
since eternity I found using frontier for everything to be very overkill. no one buys this until you give skilled people resources to let them show you, and nvidia has consistently doubled down on it. it's also the most cost/environment friendly way for downstream adopters and end users
well done 💚
