00:00There is this new paper in Nature that everyone
is talking about as a chip designer I became very
00:06curious about it myself it is a new Photonic chip
that combines two concepts ultra-high frequency
00:13radio waves and light in one single chip and
according to this paper it's up to 1,000 times
00:21faster at computing while consuming 400 times
less energy and it's actually a natural step in
00:28the evolution which started from GPUs and then
custom digital silicon and then custom analog
00:37silicon for AI and the next efficiency milestone
is likely to be photonics which allows computing
00:44with light and it's relatively new technology so
by that I don't mean that the idea is new the idea
00:51is quite mature but recently some companies have
made a great leap forward in new materials and
00:58techniques which we will discuss today photonics
computing can be tracked back to the 70s it is
01:06this idea of using light for computing why would
you want to do that well you can't be this speed
01:12of light light is the fastest thing in the
universe is the fastest possible information
01:18carrier in the complete Universe which makes
it ideal for quick computing operations and
01:24you know in electronics everything is made of
transistors with Photonic Computing you need
01:30all sort of different components like different
pathways to manipulate the light which range
01:37from wave guides to beam splitters to couplers
and the trick is these Optical components are
01:43passive and essentially drain almost no power
while in conventional AI accelerators like in
01:50GPUs you have to expand energy each time you
perform an operation you know in conventional
01:57computer chips the clock speed haven't scaled for
a while now for a decade or even more do you know
02:04why? in electronic computing the power is spent
on charging parasitic capacitances of millions
02:11and millions of logic gates each clock cycle can
you imagine that and the clock speed is limited by
02:19so-called RC time constant and that was preventing
it from scaling for a decade now but Photonic
02:27Computing overcomes both of these problems in
Photonics there's not such a thing as an parasitic
02:34capacitor or an RC constant so computations can
be performed at the speed of light at almost no
02:40power that's where this new chip comes into the
game and this chip is very interesting because
02:47it's a sort of a new flavour of an analog chip
it's an Photonic Computing engine that combines
02:54both high frequency radio waves and photonics
in one chip as you may know both radio waves
03:03and light are electromagnetic waves and they
exhibit similar wave like properties the main
03:10difference is in their frequency or in photonics
terms the wav length what's very interesting this
03:18new chip is based on lithium niobate you know
modern technology has emerged around silicon
03:25the legendary semiconductor that powers all the
laptops and data centers we have today however
03:32there is not much use of silicon in photonics
because if you look at it it's black and you
03:38cannot see through it while in photonics they
work with visible light so silicon can't do it
03:45but lithium niobate can it's actually transparent
to the visible light and some of the infrared
03:53light as well lithium niobate was discovered many
decades ago it is grown as a crystal just like
03:59silicon wafers and then sliced and polished into
many thin and transparent Wafers which are then
04:06placed on top of silicon Wafers photonics circuits
are printed on top of it lithium niobate is not a
04:13new material but with recent advancements we can
use it the same way as we use silicon we can use
04:19it to make tiny integrated photonics circuits and
that's a huge step forward using this material
04:27they've fabricated this new photonic chip which
is reportedly can make applications like wireless
04:34communication Quantum Computing and artificial
intelligence run about 1,000 times faster but
04:41before I explain how these huge performance gains
were achieved and some controversy around this
04:48work let me share with you how AI is already
making my life easier you know I write a lot
04:54of business emails and I've started to use an AI
app to help me with it the app is called TypeAI,
05:02a keyboard extension that you can use in every
app on your phone whether it's messaging apps
05:07or Outlook my favourite feature by far is the
grammar check you know I always want my emails
05:14to sound professional in the past I used to
copy paste my messages into ChatGPT to check
05:22grammar and also to rewrite them now it's much
more efficient with TypeAI because I can check
05:27the grammar in just two seconds directly where
I write it another feature I like to use is the
05:34tone changing feature for example you can type in
a reply and then in preferences choose a desired
05:41tone love it if you want to try TypeAI and support
this channel check the link in the description box
05:48below for a free trial of type AI premium version
thank you TypeAI for sponsoring this video now how
05:56does this chip work and why do we want to combine
radio waves with photonics you know if we consider
06:04mobile phones they communicate with radio waves
at a frequency like 5GHz which means 5 billion
06:13escalations per second while if we take light
light is at Terahertz range let's say 170THz so
06:21it oscillates much faster and if we compare light
to microwaves the capacity of how much information
06:28it can carry also scales similarly we can get much
more information carried by an optical signal than
06:36by an electrical signal and that's the trick in
this work they generate light by a laser in the
06:44terahertz range and this will be the information
carrier so that's where the information will be
06:50encoded and then they use a microwave signal to
actually encode this information into light let's
06:58say you have an image and you want to recognize
certain objects on it in this case they take the
07:04pixels of the image and then encode them into
a microwave signal then mix it with a light
07:11carrier after that the computations are carried
out by light passing through different optical
07:17components to put it simple different roots to
manipulate the light for example you can perform
07:24math operations for that like addition integration
or differentiation in the paper of course they
07:30claim that this chip is extremely low power and
fast and it can solve differential equations
07:37and perform image recognition tasks but the huge
advantage here that comes from mixing light with
07:43microwave signal is that this chip has Ultra broad
processing bandwidths of 67 GHz for comparison
07:51the 5G communication the state-of-the-art 5G
communication has a bandwidth of about 400
07:57megahertz which is what 200 times less which is
huge and you can actually include a lot of data
08:06in it and process it all at once you see that's
where these numbers come from.. According to this
08:12table this chip can perform tasks 1,000 times
faster than an Intel Xeon Gold processor with
08:20400 times the power efficiency so far so good
right I was very excited up until now but the
08:29closer I was getting to the end of the paper
that's where they started to slip... let's take
08:35the image detection example from the paper they
can in fact do the differentiation itself on the
08:42chip with almost no energy spent but in practice
they ignore various things here for example the
08:49power of the laser that generates light the power
of the waveform generator that generates a 67 GHz
08:56signal that consumes probably 100 Watts and then
the readout is also not addressed where they used
09:02the oscilloscope so the whole power estimates in
this paper they don't hold up I will link this
09:10paper in the description below so you can check
it out and then let me know what you think in the
09:16comments however you see combining photonics with
microwaves they can transmit at let's say 60 GHz
09:26bandwidth and with that they achieve crazy data
rates at 250 giga samples per second that's the
09:32background why they go this extra mile and combine
and mix these two technologies so as an idea it's
09:40great at the first glance I was thinking wow
they can do image recognition on this photonics
09:47chip but then I realised that the only thing
they they can actually do on this chip is the
09:52differentiation of the image and then they send it
to a conventional computer which runs the ResNet
09:59neural network you see the problem is so far they
can't do much with this chip they can't multiply
10:07matrixes with it and they also cannot reprogram
it and if you compare this work with something
10:14like Lightmatter do for example their chip is so
advanced in the case of lightmatter they're only
10:22using light for computing and their data rates are
much lower like from 100MHz to probably probably
10:29up to 1 GHz but their photonic AI chip is really
useful for AI inference they can in fact do matrix
10:38multiplication on the chip and it has been shown
to be more efficient than conventional chips so I
10:44have mixed feelings about this paper from one
point of view I'm very excited to see people
10:51mixing different domains to achieve lower power
and it can be really useful for AI applications
10:58and also Quantum Computing for example so it's
a nice try but it's still so far away from any
11:05real application that you can build and my overall
feeling is that this particular technology is more
11:13useful for the let's say data communication than
for actual Computing but we will see about that
11:21there are other companies and startups that have
some decent photonic chips and I've mentioned them
11:27in my newsletter some of them like psiquantum
and xanado are working on all Optical Quantum
11:34Computing these companies use single photons as
qubits and this approach offers huge advantages
11:42in scalability and also error correction
compared to other qubit technologies like to
11:50superconducting qubit for example we are on this
quest of finding new ways of computing and in my
11:57opinion the next efficiency milestone is likely
to be achieved with Photonics but eventually we
12:04will come to Quantum Computing as an ultimate
way when we are able to make it to work if you
12:11enjoyed this video please share it with your
friends or colleagues because this helps this
12:17channel to grow and be discovered and thank you
so much I will see you in the next episode ciao
![Welp, Here's what the Fed *Just* Said [Fed FOMC Minutes]](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fl8u23oq1E9E%2Fhqdefault.jpg&w=3840&q=75)
