📈 #47 Why software is the true driver behind faster solvers

Discover how advancements in algorithms are driving unprecedented efficiency in solving optimization problems, leaving hardware improvements in the dust.

Two weeks ago I told you why NVIDIA’s solvers might be the future.

Today I want you to do an exercise with me to see if that sentence remains true.

Suppose you were given the following choices:

1. Solve a Mixed Integer Programming (MIP) problem with today’s solution technology (modern solvers) on a machine from 1991.

2. Solve that same MIP problem with 1991 solution technology on a machine from today.

Wich option should you choose? And why would you choose that option? Are there more hardware or software advances when solving hard optimization problems?

These are the typical questions that arise when you’re facing issues solving your problems.

Especially in our current times, where we see NVIDIA pouring their GPUs for any optimization problem, and Quantum Computing claiming its advances on solving optimization problems.

So today in Feasible you’ll see:

⚙️ Why hardware alone might not be enough for future gains

📈 How software advancements are outpacing hardware in solver performance

🔮 The future of optimization - and why software innovation holds the key to real progress

In the race to optimize complex problems, should we focus on advancing hardware or on refining software?

But before we dive into it…

• This email might be too long, so your email provider might cut it. Click on ‘view entire message’ or something similar.

• Today I’m testing NotebookLM, a new AI by Google, so you can listen to this post in a podcast style. Let me know if you like it!

Now, let’s dive in… 🪂

🚧 Why hardware isn’t the golden ticket

“A problem that we foresee in the future is diminishing returns: as can be deduced from the results, having more and faster cores will not significantly improve the solvability.”

The previous sentence is not mine, it’s from a paper that discusses the progress of solvers over two decades.

So why diminishing returns happen with hardware in solvers?