When you need to enhance a technology product, one of the critical criteria is determining the direction of enhancements. What are the aspects that, when enhanced, will add value? This is also tied directly with what the end-users see as value add. But sometimes, a criterion that the end-user has been “marketed” to think as necessary is not precisely an important criterion. In fact, some of them may be non-relevant in today’s age. In my opinion, one such criterion is solver speed.

Before I explain why, one aspect that you need to keep in perspective, is that Mathematical optimization technologies are now almost 70 years old. So you can imagine that if you were a buyer in the 1990s, speed was indeed critical. With computing technology still naive, as compared to the optimization problem requirements, solver speed must have been necessary.

However, computing power is almost not a constraint now, at least for the purposes for which we need optimization solvers. A problem that may have taken days to solve in the 1990s can be solved in seconds today. While some of that can be attributed to solver performance, the primary driver is the computing power available.

So, when I see solver speed comparisons today, it makes me wonder. Why does it matter? For solving any problem, the difference between any of the leading solvers would be in minutes. And that does not matter due to how these solvers are leveraged in the real world.

No matter how much you claim, optimization solutions are, and in the near term, can not be used for true real-time planning. The farthest you can stretch is a solution that is run every half an hour or so. While for marketing purposes, you may try to label it as real-time, you know that it is not. And in that context, if a solver takes 20 minutes, and another one takes 22 minutes, the world is not going to fall apart.

Then, what exactly is the critical evaluation feature? I am a big proponent of looking for the range of problems that a solver can solve. A top-grade solver should be able to cover a wide span, including LP and MILP, convex and non-convex QP, MIQP, QCP, & MIQCP, as well as bi-linear and SOCP problems. That is most critical as we explore ways to solve problems that do not fit the conventional legacy optimization structures.

The comparison of “mine is faster” you see is primarily helpful for marketing. Or in scenarios where other features are identical. Hence, the obsession with solver speed is what I like to call the “Speed Fallacy.”

We get into this fallacy across all other technologies. We see this happening in the Generative AI race as well. Some aspects are not as critical. When applying these technologies in the real world, they become the focal point of competition. While it is okay for companies building and selling these technologies to illustrate these features as superior for marketing, it is critical for those looking to leverage these technologies to understand what is essential. It is vital not to get trapped in the “Solver Speed Fallacy.”