We always hear that future computers will use optical technology. But what will that look like for a general-purpose computer? German researchers explain it in a recent scientific paper. Although the DOC-II used optical processing, it did use some conventional electronics. The question is, how can you construct a general computer that uses only optical technology?
The paper outlines “Miller’s criteria” for practical optical logic gates. In particular, any optical scheme must provide outputs suitable for introduction to another gate’s inputs and also support fan out of one output to multiple inputs. It is also desirable that each stage does not propagate signal degradation and isolate its outputs from its inputs. The final two criteria note that practical systems don’t depend on loss for information representation since this isn’t reliable across paths, and, similarly, the gates should require high-precision adjustment to work correctly.
The paper also identifies many misconceptions about new computing devices. For example, they assert that while general-purpose desktop-class CPUs today contain billions of devices, use a minimum of 32-bits of data path, and contain RAM, this isn’t necessarily true for CPUs that use different technology. If that seems hard to believe, they make their case throughout the paper. We can’t remember the last scientific paper we read that literally posed the question, “Will it run Doom?” But this paper does actually propose this as a canonical question.
We aren’t sure we follow all of their arguments, but they do make the case that combining digital techniques with analog light representations may provide unexpected benefits. The paper indicates they’ve built a two-bit demonstration of the optical circuits and demonstrated that computation with them is possible.
What would optical CPUs mean to our community? We don’t know. While here are people who hack optics, it isn’t nearly as well-developed as what we know as a group about electronics. Still, someone will do it and, when they do, we’ll write about it.
Using optics for specific computations is nothing new, of course. You can even build a neural network with glass.
I think, for the time being, what we’ll see is optical accelerators. Devices that use an analog optical process for a single function. For instance, it is possible to do a purely optical FFT.
There are many properties of light that can be exploited this way “natively” as an analog process. The logic gate is a solid step forward but I just don’t see how that would be an optimal use of such a device; perhaps more akin to virtualization or emulation of a digital electronic computer.
Maybe later down the road we’ll get programmable accelerators; much like FPGA today. Call it the … FQOGA: Field-Adaptive Quantum-Indexed Optical Grid Accelerator (this is supposed to be tongue in cheek, but if it catches on you heard it here first!).
I question the practicality of it all. Photons are expensive in energy, whether to produce, to detect, or to amplify.
How would the megaFLOPS per watt for an optical computer compare to, say, a modern i7?
not sure where you got your numbers, but it is extremely *cheap* to produce and detect photons. there’s a reason we use optical transceivers in data centers for 100Gbps and above and not electronic wires anymore
It’s an apple-oranges comparison, but those transceivers suck well over a watt each, and still only provide one endpoint. Multiply that by the millions or billions required in a complex general-purpose all-optical computer.
Photonics promise to be up to ten million times faster while millions of times more efficient than electrons. A lot of energy is wastes as heat using electrons that would not be an issue with photons.
This all optical chip does a mind-blowing 74.8 peta-ops per second per watt.
https://www.nature.com/articles/s41586-023-06558-8
Excellent counterexample. It’s not using light to do the tasks like electrons would, and not mimicking a general-purpose computer. Entirely different paradigm. This is the direction.
What would be the clock rate ?
If I assume that an optical gate (Kerr Gate) can be toggled on and off again in 100 femtoseconds, would that mean that a clock rate of 10 THz might be possible. And since light would travel ~3 meters (~10 feet), in a vacuum and slower in all other mediums, in 100 femtoseconds, assuming some kind of clock coherence across the optical circuit, you would probably be talking about one tenth or even one hundredth of that dimension for the maximum the optical path length of the overall circuit board from input to output.
No, light would travel 30 µm in 100 femtoseconds (1e-13 s). Which still is significant distance when related to a wavelength.
And your typical everyday optical medium has you traveling at ~70% of that, so you’re at ~21 um in 100 fs.
I was wondering when I would see the delay line return to computing!
It’s about time. (sic)
https://en.wikipedia.org/wiki/Delay-line_memory
Maybe.
Cheap bare single mode fiber can support a bitrate-length product (BRLP) of a 1 gigabit/s/km. A 40 km length, costing $500 with a pair of transceivers, can support 25 Mbp/s, and can store 200 microseconds of data (5 kilobits). It also occupies a volume of roughly a liter, minimum. $100/kbit is pretty expensive memory, especially if it has a latency of up to 200 microseconds.
Standard 100 Gb grade fibers and transceivers will run you around $3000 for a pair of 100m links, that can hold 0.5 microseconds of data, 50 kilobits in each fiber, or $3000 per 100 kbits. At $30/kbit, it’s cheaper, and has much lower latency, but will still take up around 1U of rack space.
Better than acoustic delay lines, but still not ready for prime time.
And, that 100 kilobit memory will require about 10 watts of wall power, continuous.
I anticipate that the inability to achieve high enough density will ultimately prevent optical computing from becoming widespread. Visible optical wavelengths are 0.4-0.7 microns; making shorter wavelengths requires inconveniently high voltages. Modern integrated FETs have feature sizes of 0.05 microns. Packing as many optical devices in a given area as there are presently FET gates will not be possible.
The speed of light is already a limiting factor in getting signals from one side of a chip to the other. If optical devices are 10X as large, the propagation delay will also be 10X.