(Image: reve.art) Trying to keep up with technology trends is an exhilarating, if exhausting, business. I am not talking about AI here, which is hard enough, what with new advances arriving seemingly every few minutes, but rather all manner of stuff happening at other edges — where science makes its determined trek from academia to labs and code, to startups and industry, and finally into our hands, our homes and even our bloodstreams.
So having written about the frontier AI labs for months, I wanted to look upstream from the industry at this “other stuff”. This piece is about one such field where tech chatter has recently swelled from a murmur to a roar.
It is called photonics.
It’s a big deal. It is one of those technologies that has always been “around the corner”, since 1960, when the first laser was built. And when in 1966 a fellow named Charles Kao showed that light could be used to move information, the race was on (he won the Nobel Prize for that, and is now addressed as Sir Charles Kao).
We’ll get back to photonics in a moment, but a little history is necessary. In 1965 Intel’s Gordon Moore famously made a simple projection — that the number of transistors on a chip would double about every 18 months or so. It was so prescient that for six decades we became accustomed to computers getting smaller, faster and cheaper almost automatically.
Entire industries were built on the assumption that next year’s chip would outperform this year’s by a comfortable margin. It made computing a million times cheaper in mere decades, and turned science fiction into consumer electronics. The personal computer, the laptop, the internet, mobile phones, GPS, streaming media and now artificial intelligence all owe their existence to the extraordinary persistence of what has come to be known as Moore’s Law. It may prove to be one of the most economically consequential (and reliable) industrial trends in human history.
Moore’s Law is now sunsetting, the victim of the limits of physics.
The critical features of today’s most advanced transistors are now measured in atoms — the insulating layer between transistors is four or five atomic layers thick, the silicon channels a couple of dozen atoms wide. At those scales, electrons stop behaving obediently and start behaving like delinquents — tunnelling through barriers they are supposed to respect, leaking current, generating heat. And the microscopic copper wiring that connects billions of transistors has its own crisis. When it is squeezed to a dozen nanometres, its resistance soars. This means that the wires (rather than the transistors they connect) increasingly dictate how fast a chip can run and how hot it gets.
Chipmakers have responded with ambitious workarounds, but the underlying engine of decadeslong exponential improvement driven by Moore’s Law has slowed. The industry has known this for a decade. What it has lacked is a successor.
Enter photonics — communicating and computing with light instead of electricity — which after 60 years of R&D gestation has finally (in the last 18 months or so) broken out of the laboratory and into the product catalogues of the world’s most important technology companies.
Unparalleled parallelism
Why is light better than electrons? Photons are massless, generate essentially no heat in transit, and do not interfere with one another. Better still, dozens of different colours of light can travel down the same waveguide simultaneously, each carrying its own stream of data — parallelism that copper cannot achieve. Furthermore, we now understand light as a communications medium. Data centres already use fibre-optics between buildings and racks (as do we, between our home laptops and the street outside).
Now it’s coming into the chip.
The numbers are startling. Optical interconnects can move data between chips up to 100 times faster than their electrical equivalents. Nvidia’s new co-packaged optics switches — shipping this year, built with the Taiwan Semiconductor Manufacturing Company — claim 3.5 to 5 times better power efficiency and tenfold improvements in network resilience.
Lightmatter, a photonics startup now valued north of $4-billion, has demonstrated photonic processors running AI workloads roughly 10 times faster than conventional GPUs. Furthermore, there is a roadmap – by adding those additional colours of light — that its founder claims extends speed increases beyond a hundredfold. Even discounting for the traditional exuberance of startup CEOs, these are not incremental gains. They are the kind of leap Moore’s Law used to deliver over a decade, arriving in one single product generation.
And beyond AI? Lidar (the laser radar that lets autonomous vehicles see) is being shrunk from a spinning rooftop gizmo on a car onto a fingernail of silicon. Photonic sensors can detect biomarkers in blood, perform medical spectroscopy, and may end up in wearables monitoring glucose without needles. Telecommunications, quantum computing (several leading contenders use photons as their qubits), navigation-grade gyroscopes, industrial sensing, environmental monitoring — all are downstream of cheap, mass-produced photonic silicon.
When a component drops from laboratory instrument to commodity chip, backed by the credibility of a company like Nvidia, entire industries reorganise around it. This is what happened in the 1970s with the integrated circuit. We have been here before.
Notwithstanding all the hype around this, there are plenty of reasons for caution, as is to be expected for all new paradigm-jumping technologies. Photonics is great for communications — it is not yet quite there for actual computation. It is fiddly and complex to control (silicon and light get pissy with each other, so exotic materials are needed).
Also, the semiconductor industry moves very tentatively because every manufacturing decision involves investments measured in tens of billions of dollars. But once companies as dominant as Nvidia start actually shipping photonics-based product, the caution-dyke has already been breached.
The technology that was forever 10 years away is now roughly zero years away, and it is arriving precisely when Moore’s Law has run out of road. Moore’s Law is dead, but, as always in technology progress, there is always something else, something new, something better, something transformational. DM