CertiSum Achieves Bitwise-Identical MXFP4 Training Across GPUs
Reproducibility at 4-bit precision removes a practical blocker for teams that must audit or verify training runs, since bitwise-identical states make cross-vendor and cross-layout debugging tractable. Engineers running multi-vendor fleets mixing AMD and NVIDIA hardware gain a way to compare runs without engineering tolerances. Teams choosing between BF16 and MXFP4 can weigh the 11.8–60.5% latency overhead against the retained speedups over BF16 baselines.