"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.