zardinality

"For extreme performance, we discover and use a behavior-out-of-doc PTX instruction: ld.global.nc.L1::no_allocate.L2::256B. This instruction will lead to an undefined behavior: accessing volatile GPU memory with non-coherent read-only PTX modifiers .nc. But the correctness is tested to be guaranteed with .L1::no_allocate on Hopper architectures, and performance will be much better. If you find kernels not working on some other platforms, you may add DISABLE_AGGRESSIVE_PTX_INSTRS=1 to setup.py and disable this, or file an issue."

In the introduction of the paper it says: "Despite its excellent performance, DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training. In addition, its training process is remarkably stable. Throughout the entire training process, we did not experience any irrecoverable loss spikes or perform any rollbacks." They have indeed a very strong infra team.

I completely agree with the point you made, as in the choice is much more nuanced, and you do not have to guess my position by the punchline. I never thought of US, in fact I do not know the vaccine policy there. If I was thinking some country while I wrote the comments that would be EU from 2021, when people have to endured some months of inconvenience for not taking any shot. I know some anti-vaxxers even bought virus from other country so that they did not need the vax pass.

Yeah, it is like the summer that year in EU. In summer EU still has tens of cases per day, but it was a quite small number in retrospective angle.