WEBVTT

00:00:00.000 --> 00:00:05.200
When you think of light traveling down some kind of narrow channel, which I know you do all the time,

00:00:05.200 --> 00:00:09.440
you probably think of high-speed internet connections, telephone lines, or even those

00:00:09.440 --> 00:00:13.840
pre-lit Christmas trees. But did you know that optical connections are actually taking the

00:00:13.840 --> 00:00:19.600
place of traditional metal traces in actual computer chips? This field is called silicon

00:00:19.600 --> 00:00:25.200
photonics, and it's not some kind of gimmicky speculative technology. Products that use silicon

00:00:25.200 --> 00:00:29.760
photonics are already being deployed in data centers by the millions to help keep up with

00:00:29.760 --> 00:00:34.720
how much pressure we're increasingly putting on cloud systems these days, especially considering

00:00:34.720 --> 00:00:41.520
how rapidly AI is expanding. The basic idea is to use photons instead of electrons to move data

00:00:41.520 --> 00:00:47.280
around. The best part is that chip makers are manufacturing the means to do so directly onto

00:00:47.280 --> 00:00:53.040
silicon wafers, meaning that you don't need costly exotic materials in order to pull this off.

00:00:53.040 --> 00:00:59.440
Silicon photonics uses waveguides that the light can move through, as well as freaking laser beams,

00:00:59.440 --> 00:01:05.840
as a light source. These hybrid silicon lasers are made from both silicon and other semiconductors,

00:01:05.840 --> 00:01:11.120
such as indium and gallium compounds, so they can be built directly onto the chip,

00:01:11.120 --> 00:01:16.480
obviating the need for extra laser equipment outside the chip, which can be costly.

00:01:16.480 --> 00:01:21.600
This technology has already been in use for a few years in transceivers that convert electrical

00:01:21.600 --> 00:01:27.280
signals to and from optical ones. You plug them into a server and use fiber optic cable to connect

00:01:27.360 --> 00:01:32.400
to another server elsewhere in a data center, up to several kilometers away. A piece of tech

00:01:32.400 --> 00:01:37.280
that's also proven useful for supporting current 5G networks. But there's been a more recent

00:01:37.280 --> 00:01:42.320
development that might make those external transceivers obsolete. Instead of relying on a

00:01:42.320 --> 00:01:47.920
separate transceiver to handle the actual photonics part of the chain, recent work has looked more at

00:01:47.920 --> 00:01:54.400
integrating silicon photonics directly onto a chip package. In 2022, Intel showed off a laser

00:01:54.400 --> 00:02:01.360
setup that used tiny micro rings to save enough space to fit the actual photonics next to the die,

00:02:01.360 --> 00:02:06.320
with each ring producing different light wavelengths, allowing more data to be transferred

00:02:06.320 --> 00:02:11.200
than if you just used a single wavelength. Intel also revealed a chip they had been working on for

00:02:11.200 --> 00:02:18.000
DARPA in 2023 that used optical chiplets. These chiplets sit on the same package as the actual

00:02:18.000 --> 00:02:22.880
compute die and are responsible for doing the conversions between electrical and optical.

00:02:22.880 --> 00:02:26.720
So you'd be connecting fiber optics directly to these chiplets,

00:02:26.720 --> 00:02:31.200
again eliminating the need for external transceivers when this tech hits the marketplace.

00:02:31.200 --> 00:02:36.240
But beyond the cost advantages of moving photonics as close to the compute die as possible,

00:02:36.800 --> 00:02:42.800
we might need silicon photonics more and more as the years go by. As our bandwidth needs just

00:02:42.800 --> 00:02:47.200
continue to grow, there's a good chance we'll start hitting the limits of what's possible

00:02:47.200 --> 00:02:52.640
with standard metal traces in terms of power or data integrity. The thinking is that it could be

00:02:52.720 --> 00:02:58.000
cheaper to use silicon photonics instead of trying to find ways to shove more and more

00:02:58.000 --> 00:03:03.840
data down a metal trace, and that photonics could offer a superior error rate to traditional

00:03:03.840 --> 00:03:09.360
electronics. But silicon photonics isn't just about ensuring the cloud can continue to deliver

00:03:09.360 --> 00:03:15.120
fine YouTube videos like this one to your screen. Because the general idea is put optical technology

00:03:15.120 --> 00:03:20.320
into silicon, it has other applications such as LiDAR for self-driving cars, healthcare

00:03:20.320 --> 00:03:25.760
diagnostics that involve shooting light into tissues, and even better sensors for AR and VR

00:03:25.760 --> 00:03:32.240
headsets. It's unclear if silicon photonics will have a big role to play in home PCs at some point,

00:03:32.240 --> 00:03:36.400
but if you really want optical tech in your personal rig in the meantime,

00:03:36.400 --> 00:03:40.080
you could always just put in a Blu-ray drive. They're not dead yet.

00:03:40.080 --> 00:03:44.960
And thank you for popping by this episode of TechWiki. Like the video if you like it,

00:03:44.960 --> 00:03:51.760
and dislike it if you dislike it, check out our other videos, comment below with video suggestions, and don't forget to subscribe and follow. Did you get all that? Okay.