[ahzhou]

I might be missing something, but DeepSeek’s recipe is right there in plain sight. Most of the cost efficiency of DeepSeek v3 seem to be attributable to MoE and FP8 training. DeepSeek R1s improvements are from GRPO-based RL.

Interesting to note - we have no idea how much R1 cost to train. To speculate - maybe DeepSeek’s release made an upcoming Llama release moot in comparison.

[pptr]

What is different about Deepseek's use of MoE vs all the other MoE models that makes training more efficient?

FP8 training and GRPO make sense to me, but that only gets you a 4x improvement total, right?

[ahzhou]

They slightly restructure their MoE [1], but I think the main difference is that other big models (e.g Llama 504B) are dense and have higher FLOP requirements. MoE should represent a ~5x improvement. FP8 should be about a ~2x improvement.

We don’t know how much of a speed improvement GRPO represents. They didn’t say how many GPU hours went into to RLing DeepSeek-r1 and we don’t have a o1 numbers to compare.

There’s definitely lots of misinformation spreading though. The $5.5m number refers to Deepseek-v3, not Deepseek-r1. I don't want to take away from HighFlyer's accomplishment, though. I think a lot of these innovations were forced to work around H800 networking limitations, and it's impressive what they've done.

[1] https://arxiv.org/abs/2401.06066

[karmakaze]

It's interesting that only having access to less powerful hardware motivated/necessitated more efficient training--like how tariffs can backfire if left in place too long.