Time-to-target
Platform-specificD-Wave's relaxation of time-to-solution that times solvers to a target energy (originally set by quantiles of a reference annealer's short-run samples) rather than to the exact ground state.
Time-to-target (TTT) is D-Wave’s deliberate relaxation of time-to-solution: on hard instances an annealer may almost never sample the exact ground state, so King, Yarkoni, Nevisi, Hilton and McGeoch (2015) instead timed solvers to a target energy that the annealer itself sets. Strictly a named metric with a target-setting procedure rather than a standalone scored benchmark, it is filed here alongside the annealing benchmarks it serves.
How it works
The 2015 protocol lets the quantum annealer set the bar. Sample a reference annealer (a D-Wave 2X in the study) for a short fixed time (15 to 352 ms) and place the target energy at a low quantile (q = 0.001, 0.01, or 0.1) of the sampled energy distribution. Competing solvers then race to produce a solution at or below that target, reported as the expected time to reach it: TTT_total (wall clock including programming and readout) or TTT_anneal (anneal time only), computed from the companion samples-to-target (STT) count. Modern usage is looser: typically TTT(T) = T · log(0.01) / log(1 − P(T)) for a chosen target energy, i.e. the TTS formula with the ground-state requirement swapped out.
Strengths and limitations
TTT rewards what annealers are actually good at (fast, decent-quality samples), and the metric remains in active use in Ising-machine benchmarking, from a 2023 sparse-Ising optimization study to a 2026 analysis of classical adiabatic annealing. The caveats are real, though. The introducing paper is vendor-authored (five D-Wave employees benchmarking D-Wave’s own 2X against single-threaded CPU software), and its headline 2–15x total-time (8–600x anneal-only) results come with an explicit disclaimer of any quantum-speedup claim; the paper was never journal-published. Deriving targets from a reference annealer’s samples also bakes that machine’s strengths into the test. And the name has drifted: later papers often drop the quantile procedure (and the King et al. citation), so a reported “TTT” is not always this protocol; nor is it the older classical time-to-target plots of Aiex, Resende and Ribeiro (2002), which share only the name.