Nice result, though the number I would actually want to see is what fraction of a step your uplink can sustain. Contribution in these setups is rarely linear in FLOPs, it gets clipped by however fast you can push gradients back out. A fast card behind a slow line contributes about as much as a slow card.

congrats you built a space heater with a leaderboard

泼一点冷水，但不是唱衰。家用宽带的上行带宽是这类方案真正的瓶颈，同步一次梯度的开销往往比本地算一步还高，所以实际有效贡献要打不少折扣，值得自己算一笔账，别只看跑了多少小时。 不过我认为参与门槛降下来这件事本身是有意义的。过去想理解大规模训练怎么调度、怎么容错，只能看论文，现在能亲手挂一台机器看它怎么被排进去、掉线之后又怎么被补上，这个认知上的收益比那点算力大得多。

电费是真心疼 😭 但学到东西了 值 我准备再挂一台 😤🔥