for our next open source project, would it be more useful to do an o3-mini level model that is pretty small but still needs to run on GPUs, or the best phone-sized model we can do?

NVIDIA Statement:

We have identified a rare issue affecting less than 0.5% (half a percent) of GeForce RTX 5090 / 5090D and 5070 Ti GPUs which have one fewer ROP than specified. The average graphical performance impact is 4%, with no impact on AI and Compute…

@deskflag 没有任何别的解释能说得通。如果是生产过程中没有筛出来缺陷品，那么在用户手里应该表现为玩游戏崩溃不稳定或者花屏等等故障。但如果NVAPI/驱动软件层面能识别，说明封测时已经发现缺陷并且作了对应的屏蔽、降级处理。

除了闲鱼和一些渠道上还在卖本地部署大模型教程外，绝大多数社媒上，我观察到，本地部署的热度终于下去一点了。

想来是很多人自己试过后，终于明白本地量化后的模型是残废是什么意思了。

跑工作流的人可能也发现了，自己那张卡跑一个并发还行，上了两个就明显变慢，几乎不可用。