How much shorter is the route? Reading D-Wave’s 200-fold claim
D-Wave gave two versions of its Stride software 15 seconds to plan a route. Its headline number measures how close each answer came to the best possible tour.
Reading tools for text size, copying, printing and your reading list
Give a route planner 15 seconds to visit every stop on a list and get back to the depot. When the clock runs out, you can judge its answer by how much farther the driver would have to go than on the shortest possible tour. That extra distance is what makes D-Wave’s November 4 benchmark interesting. Its Stride software was competing against an earlier version of itself.[1][3]
A number about the route
In the company’s tests covering up to 10,000 locations, the new version finished an average of 0.02% above the optimum with 15 seconds of solver time. D-Wave says the remaining gap was more than 200 times smaller on average than with the previous version at the same runtime. A separate solver or algorithm had already established the reference optima, according to the company’s footnote.[1]
Imagine a delivery circuit whose shortest possible tour is 100 miles. One planner returns a 110-mile tour and another returns a 101-mile tour. The extra mileage fell from ten miles to one, a tenfold improvement in the gap. The driver saves nine miles. Both planners could have run for exactly the same amount of time. These distances are an example to explain the arithmetic.[3]
How to read the company benchmark
What does “better” measure?
Read the two tests separately. D-Wave used the earlier version of Stride as the comparison in both.
| Task | Time budget | Answer quality |
|---|---|---|
| Plan a tour | Both versions get the same solver time for a given group of routes. | Measure how much longer each route is than the shortest possible tour. A smaller gap wins. |
| Balance an assembly line | The newer version gets a shorter run. | Compare the assignments returned at the two time limits. This test asks whether the updated solver found a better assignment sooner. |
The route claim compares gaps in answer quality. To compare speed, both versions would have to reach the same quality target and report the time it took. Read D-Wave’s benchmark announcement and footnote.
The average also hides something a dispatcher might care about. A planner could do very well on most routes and leave one awkward route much longer than it needs to be. A per-route list of results would let a reader spot that pattern. The company’s release groups the tests by size and gives averages for each group.[1]
Two clocks, two different tests
D-Wave also tested how software assigns tasks to stations on a simple assembly line. Here the two versions did get different time budgets. According to the company, the new version found a better assignment in 15 seconds than the preceding version found in 120. A factory manager would still need to test a schedule against the plant’s own constraints before using it on the floor.[1]
The timing of the announcement matters, too. D-Wave’s Leap release notes had already recorded improved Stride performance on some problems on October 21. November 4 brought the public benchmark figures and a three-course training program for developers. Stride combines conventional and quantum computing, and D-Wave attributes the update to changes in its algorithm.[1][2]
What would a dispatcher want to know?
A dispatcher choosing software could give several planners the same stops, the same time window and the same rules for a valid delivery. They could then compare the distance each planner returned and the total wait from submitting the stops to receiving the route.[1][3]
D-Wave’s published comparison leaves those purchasing questions open. It gives no end-to-end timing or same-budget test against a leading conventional route planner. It also leaves open how much each part of the hybrid system contributed to the improvement.[1]