A new microchip technology capable of optically transferring data could solve a severe bottleneck in current devices by speeding data transfer and reducing energy consumption by orders of magnitude, according to an article published in the April 19, 2018 issue of Nature.
Researchers from Boston University, Massachusetts Institute of Technology, the University of California Berkeley and University of Colorado Boulder have developed a method to fabricate silicon chips that can communicate with light and are no more expensive than current chip technology.
The electrical signaling bottleneck between current microelectronic chips has left light communication as one of the only options left for further technological progress. The traditional method of data transfer–electrical wires–has a limit on how fast and how far it can transfer data. It also uses a lot of power and generates heat. With the relentless demand for higher performance and lower power in electronics, these limits have been reached. But with this new development, that bottleneck can be solved.
"Instead of a single wire carrying around 10 gigabits per second, you can have a single optical fiber carrying 10 to 20 terabits per second—so a thousand times more in the same footprint," says Assistant Professor Milos Popovic (ECE), one of the principal investigators of the study, whose team was previously at University of Colorado Boulder where part of the work was done.
"If you replace a wire with an optical fiber, there are two ways you win," he says. "First, with light, you can send data at much higher frequencies without significant loss of energy as there is with copper wiring. Second, with optics, you can use many different colors of light in one fiber and each one can carry a data channel. The fibers can also be packed more closely together than copper wires can without crosstalk."
Source: http://www.bu.edu/eng/2018/04/18/a-new-era-of-microelectronics/
(Score: 5, Interesting) by bob_super on Monday April 23 2018, @05:41PM (3 children)
People have been talking about this for decades.
It turns out it's pretty hard to displace the convenience of the PCB. Optical attachments of all kinds don't print the same way, don't go through reflow the same way, and typically get in the way of heatsinks...
It's a bit like GaAs, SiGe and other exotic materials: We know they're better, but the Si guys keep upping their game and not leaving enough advantage in price/performance/thermal for the exotic stuff to become mainstream. Why bother manufacturing with fibers, when it's not that hard to get multiple 25G x16 chip-to-chip on a PCB? Once you need the layers for the twelve supplies already, crosstalk isn't typically a big deal. And getting a chip to output 10 Tb is a problem of power, not PCB traces.
It will come, maybe at 100G/lane or above, when fiber weave become a real pain. Just kidding, that will again get pushed an order of magnitude back with QAM and other analog modulations, which the Si guys already have at 50G today.
(Score: 3, Interesting) by fyngyrz on Monday April 23 2018, @06:17PM (2 children)
Is this about PCBs? Or is this is more about chips?
First line in TFS, emphasis mine:
So...
(Score: 3, Informative) by bob_super on Monday April 23 2018, @07:15PM (1 child)
Third paragraph:
Granted, they could be talking about going between chips on an Si interposer, but that's essentially a different kind of PCB. But distance isn't really an issue until you go 20 inches and/of a few connectors away.
(Score: 3, Insightful) by fyngyrz on Monday April 23 2018, @09:39PM
RF signals in conductors present an entirely different set of challenges as compared to optical paths; we're up against those limits right now, only incrementally nudging them here and there. CPU speeds are not increasing very quickly at this point, and adding 2D area to a CPU chip – which deals with internal distances far less than 20" – tend to not speed it up... quite the contrary. Optical rates and signal modulation densities are far, far in excess of what we can do in what is effectively a wire, be it an oxide path or conventional copper, etc., and if - for instance - an output bit in a CPU could talk to a bit in something else much faster without dealing with a significant inductive load (the wire and its coupled impedances), that would be a significant gain, one worth a whole waxload of candles. Likewise, if input got there at Terahertz rates instead of gigahertz rates... that's a serious win.
If. :)
(Score: 2) by inertnet on Monday April 23 2018, @08:34PM (1 child)
If you still need traditional logic gates on both ends, you'll also need a conversion on either end.
I understand the benefits for data transport, but I don't see how this would speed up data processing.
(Score: 3, Interesting) by fyngyrz on Monday April 23 2018, @09:44PM
Many limitations on CPU speed are consequences of waiting for every signal to arrive and settle before the next step in the process can be taken. Transport of signals within the chip, therefore, are a good target for an overall speedup of data processing. Likewise, getting data in and out of memory is a severe bottleneck; if it can be worked around, things speed up overall.
Basically, data transport is a bottleneck in many physical designs, large and small. Significantly opening it up would be a win for how soon we get the results we desire.