QF8 Hardware Cost: 3.5× Cheaper Multiplier Than FP8

The QF8 multiply path (7-bit CLA adder + sign XOR + control) costs ~66 gates vs. ~231 for FP8 mantissa multiplication. The full QF8-Medium MAC unit (814 gates, 0.038 pJ/MAC at 7nm) achieves near-INT8 area/power while providing floating-point dynamic range. A 128×128 systolic array achieves 50.5 TOPS/W vs. FP8's 27.6 TOPS/W. Log-domain break-even at ~7–8 bits.