Why, After All These Years, MZI-Based Transistorlessness Might Finally Be Here

It has been reported that Mach–Zehnder interferometer (MZI)–based, transistorless photonic computing may be approaching practicality — and yes, that makes many in the field both excited and deeply skeptical. Optical computing has been the perennial “five years away” story since the 1980s. The physics has always been beautiful; the problem has been making it survive the real world. So what’s different now? Allegedly, three trends are colliding in a helpful way: AI’s relaxed precision needs, smartphone-grade thermal tricks, and a quieter engineering pivot on the MZI’s oldest weakness.

Why MZIs falter

Here’s the rub. MZIs do their work by splitting light, sending it down two paths, and recombining it — interference does the math. But silicon’s refractive index changes with temperature, and that tiny change shifts interference patterns. A degree or two drift and your output looks wrong. In the lab? Fine. In a rack of hot chips? Not so much. Historically, keeping MZI meshes thermally stable required active heating, calibration, and power-hungry feedback that ate into the energy gains photonics promised. No wonder the transistor has stubbornly refused to die.

What’s changed

The first wrinkle in the old story is AI. Modern inference can work at 8-bit, even 4-bit, precision. Neural nets tolerate sloppier arithmetic than old-school HPC codes — which means MZIs can tolerate more thermal drift before results go off a cliff. Next, thermal-management lessons from phones — better packaging, heat spreaders, low-power control loops and aggressive local throttling — are migrating into datacenter and chip design. And engineers aren’t standing still: it has been reported that teams are combining athermal device design, smarter calibration algorithms, and closed-loop control to blunt the MZI’s sensitivity.

Why it matters

If these pieces actually fit together, the emotional moment is obvious: relief mixed with a dash of vindication. Photonics wouldn’t have to be perfect; it only needs to be good enough and far more efficient at the precision bounds AI now accepts. Still, caveats loom. Scaling, yield, integration with existing silicon fabs, and the long tail of real-world thermal chaos remain hard. Could MZIs finally help topple the transistor as computing’s workhorse? Maybe. Or maybe this is another detour on the long road light has been traveling toward the CPU. Either way, it’s worth watching — and for once, not with the weary sigh of someone who’s heard this promise a dozen times before.

Sources: write.as, Hacker News