FLOPS & Precision Formats ☆

FLOPS count floating-point operations per second; AI hardware quotes them per precision (FP64 → FP4), roughly doubling each halving of bits, and vendors often quote sparsity-doubled figures — always check the footnote (records here label dense vs sparse). Training compute totals (10²⁵–10²⁶ FLOP for frontier runs) anchor scaling laws and even export regulation thresholds.