Kummer Strikes Back: New DH Speed Records

This paper introduces high-security constant-time variable-base-point Diffie–Hellman software using just 274593 Cortex-A8 cycles, 91460 Sandy Bridge cycles, 90896 Ivy Bridge cycles, or 72220 Haswell cycles. The only higher speed appearing in the literature for any of these platforms is a claim of 60000 Haswell cycles for unpublished software performing arithmetic on a binary elliptic curve. The new speeds rely on a synergy between (1) state-of-the-art formulas for genus-2 hyperelliptic curves and (2) a modern trend towards vectorization in CPUs. The paper introduces several new techniques for efficient vectorization of Kummer-surface computations.