📈 #85 Where did the time go? Part V: hardware (Part B)

Does faster computation mean faster optimization?

Last week we explored why optimization has remained loyal to CPUs while AI migrated to GPUs and quantum startups promise revolutionary speedups.

We saw that CPUs are accidentally perfect for optimization's irregular branching patterns and discrete decisions.

But CPUs are hitting three fundamental walls: memory bandwidth, instruction-level parallelism, and power limits.

So what alternatives actually exist? And more importantly: should you care?

Let’s explore that in today’s Feasible article:

• 🏎️ What faster hardware is on the shelf?

• 🎯 Extra computational power: what can we still do?

Remember this is the fifth part of a 6-part series of posts focused on tractability. We started with bencharking before fixing, then large-scale problems, three more parts centered on symmetries (what they mean and classical techniques, more advanced techniques, and ML-guided techniques and practical ways to dealing with symmetries), and finally algorithms and parallelization.

Ready? Let’s dive in… 🪂

🏎️ What faster hardware is on the shelf?

Yes, we’ve been stuck with CPUs for some reasons, but that doesn’t mean there are no alternatives at all to try boosting your computations.

In fact, there are plenty of them.

🇯🇵 Digital Annealer and the like

I’d like to highlight this as the first option as I already worked with this hardware.

It was February 2019 and I started working for Fujitsu. I was motivated by trying new quantum-inspired hardware, the project at a telecom company, and the team.

In just two months of Proof of Concept, we saw the real advantages: a process that took more than 2 hours of running time was solved with similar results in just…