00:00:01
analog computers once dominated the world but
they were complex noisy and inaccurate so in the
00:00:10
early 60s we shifted to digital chips precise
deterministic and powerful but this technology
00:00:19
is now reaching its physical limits and we are
actually at the brink of yet another paradigm
00:00:25
shift in computing probabilistic computing or
some call it thermodynamic computing and this new
00:00:34
technology completely flips the script instead
of fighting the noise which we've done for the
00:00:41
past 60 years we are now embracing the noise using
it as a computational resource and this approach
00:00:50
reportedly allows for 100 million times more
energy efficient computing compared to the best
00:00:58
NVIDIA GPUs as a hardware engineer hearing these
orders of magnitude make me of course skeptical
00:01:06
at first it seems they've broken something but now
it all makes sense let me explain modern classical
00:01:14
computers are built out of transistors that are
deterministic very precise objects they operate
00:01:22
by switching between two states 0 and 1 and
this binary system powers nearly all modern computational
00:01:29
tasks however the world around us
is not binary it's governed by probabilistic
00:01:36
rules for example when you want to find the best
solution out of a large number of solutions like
00:01:43
finding the best route for an Uber or predicting
weather or financial markets probabilistic
00:01:50
algorithms are far more effective and these are
very hard and expensive to do on a classical
00:01:58
digital computers also all generative AI tasks
are probabilistic distributions at the moment we
00:02:06
are trying to solve this probabilistic problems on
classical computers by forcing a digital computer
00:02:13
to behave as a probabilistic system and to do
that to simulate this indeterminism requires
00:02:20
a large number of transistors and this is very
energy draining in addition to being too slow
00:02:28
already in 1982 Richard Feynman suggested that
rather than forcing traditional computers which
00:02:35
are deterministic to simulate the probabilistic
nature we need to build a new kind of computer
00:02:41
which itself is probabilistic in which the output
is not a unique function of the input so about 40
00:02:50
years later researchers from MIT the University
of California Santa Barbara Stanford and startups
00:02:58
like Normal Computing and Extropic all working
on building probabilistic computers probabilistic
00:03:08
computing is a very interesting concept first of
all it brings more uncertainty to our lives like
00:03:15
we don't have enough already but as you will see
it sort of bridges the gap between classical and
00:03:23
quantum computing because it's able to address a
subset of quantum problems but at the same same
00:03:30
time working at room temperature and so this tiny
devices that we will discuss just now they harness
00:03:38
the noise from environment and amplify it and
this is just mind-blowing to me because with my
00:03:45
experience in first analog and then digital design
we used all different techniques to fight noise to
00:03:52
get rid of noise and here we are harnessing it
we embracing it and this actually flips my whole
00:03:59
world around so let me explain how it works the
foundation of the probabilistic computing is the
00:04:07
Boltzmann law which describes how particles such
as atoms and molecules distribute themselves in a
00:04:16
system essentially this law states that particles
are more likely to occupy lower energy states than
00:04:24
higher energy ones it turns out we can harness
this law to find the most most probable state
00:04:31
for a given system and the answer is found in
equilibrium the system essentially searches
00:04:38
through many different possible configurations
similar to how molecules of gas in a box moving
00:04:44
around until they reach the state of equilibrium
if you don't understand how probabilistic computer
00:04:52
works just yet make sure to stay till the end
of the video and you will now we all know the
00:04:59
foundation of all modern classical computers
bits zeros and ones which CPUs and GPUs use
00:05:07
to perform logical operations now on the other
side of the spectrum is qubit these are quantum
00:05:14
bits which exist in a superposition of 0 and
1 simultaneously and these are the foundation
00:05:21
of quantum computing in the middle of the
spectrum is a p-bit so-called probabilistic bit
00:05:29
which is designed to naturally fluctuate between
two states 0 and 1 in a purely classical non
00:05:37
quanum manner and this oscillation is happening due
to its thermal energy now the probabilistic bit
00:05:45
and quantum bit are very different but they
share some similarities you know the p-bit
00:05:50
isn't quantum mechanical but it isn't completely
classical either because it's this there's this
00:05:56
true randomness and fluctuations built into
it right so the way the p-bit works is it's
00:06:02
like a coin flip but it changes the probability
sometimes it will give you all heads sometimes it
00:06:08
will give you 50/50 sometimes it will give you
all tails so this tunability is very important
00:06:13
because when you connect them you want the system
to go somewhere here just like quantum bits which
00:06:21
can be built in a different ways p-bit can be
built in different technologies typically we
00:06:27
are talking about CMOS plus something CMOS plus
super conductivity CMOS plus magnetism so-called
00:06:35
MGTs and this magnetic memory cells have a
remarkable property of gathering noise from
00:06:42
environment and very well amplifying this noise so
they are perfectly naturally unstable from the
00:06:49
very beginning and this allows them to naturally
fluctuate between 1 and -1 and this is
00:06:57
exactly the property we are looking for for a
perfect tunable p-bit now how does a probabilistic
00:07:06
computer work here we get a hint from nature from
physics just like molecules of gas interact in a
00:07:16
box or how our brain works our brain is probably
the most familiar to you probabilistic machine
00:07:25
which runs 100 trillion parameters neural network
on just 20W of power this is remarkable and we
00:07:32
would like to get inspiration from here the way
this computer works is it's you take p-bits and
00:07:39
then you make them you interconnect them so it
isn't very different from what people might be
00:07:45
familiar from the context of neural networks
you could view the network of the p-bit as some
00:07:51
kind of a neural network but with usually our
neural networks are feed forward you know it
00:07:57
goes from left to right like our digital search
but with p-bits the networks are typically for
00:08:03
most applications there there's feedback I talk to
Anastasiia Anastasiia talks to me and then this this
00:08:10
network as a whole goes somewhere right it evolves
it evolves as a function of time a p-bit takes an
00:08:19
input from other p-bits and creates a weighted sum
which is so common in machine learning based on
00:08:26
the weighted sum we get the probability at the
output derived from the inputs average and that's
00:08:34
very important property which makes them very
well-suited for probabilistic algorithms AI and
00:08:40
machine learning tasks what's so interesting this
computing principle is very similar to the this
00:08:48
year winners of Nobel Prize in physics which
got it for the Boltzmann machine mathematically
00:08:55
the p-bits model is exactly like this Boltzmann
machine a neural network that embraces chaos to
00:09:02
solve optimization problems let me know what you
think about the probabilistic computing in the
00:09:08
comments I love reading your comments and if you
enjoying this video subscribe to the channel not to
00:09:15
miss the future updates for you it costs nothing
but for me it means the world now let's dive into
00:09:23
thermodynamic computing it's a new flavor of
probabilistic computing where where we have
00:09:30
a noisy system that basically harness noise
as a computational resource to solve problems
00:09:38
this computing approach has been gaining a lot
of momentum recently you may have heard about
00:09:44
startups Normal Computing and Extropic building a
thermodynamic computer and many of you messaged me
00:09:51
asking to make a deep dive into this technology so
enjoy Extropic is a startup building a thermodynamic
00:09:59
computer and creating energy-based models
that perform computation through heat dissipation
00:10:08
as you may have guessed from the name they are
using the second law of thermodynamics which
00:10:13
states that in a natural process the total entropy
of an isolated system always increases as we just
00:10:22
discussed before they're using system natural
tendency to minimize energy as a computer resource
00:10:30
to perform computation as a physical process so
this is very similar to what we do with
00:10:37
probabilistic systems and like you said you know
one mode of working is you make the solution of
00:10:43
your problem the equilibrium of some statistical
system so you started in some non-equilibrium
00:10:50
state as it evolves in time it equilibrates but
the algorithm designer was clever such that the
00:10:57
equilibrium state is the answer to a problem that
you wanted to solve in practice Extropic is
00:11:04
using Josephson junctions or so-called JJ's to build
probabilistic bits that are very fast and these
00:11:12
JJ's consist of two superconductors separated by a
thin insulating layer and when the energy barrier
00:11:20
is low enough we get fluctuations so with this
JJ's they create a lot of fast probabilistic bits
00:11:29
they configure it and let it run and it's
probabilistically explore different states and the
00:11:35
resulting thermodynamic equilibrium is actually
the solution to the problem it appears that this
00:11:42
computing approach has a whole set of advantages
for many modern AI computing tasks for example for
00:11:50
diffusion based models like DALL-E 2 which is used
to generate images by first noising it and then
00:11:57
denoising it and as you may know on a classical
computer it takes a while you have to wait for it
00:12:03
while according to Extropic they can do it way
faster on a thermodynamic computer they can even
00:12:11
run transformer models on it Extropic estimates
that transformers on a thermodynamic computer are
00:12:19
up to 100 million times more energy efficient
than on a GPU Cloud now understanding this under
00:12:29
understanding the potential of technology and all
involved technical risks is very challenging that
00:12:35
what I saw working with investors and executives
over the last year advising them on semiconductor
00:12:41
and AI technologies and now I'm happy to share
with you the whole strategy in my new course
00:12:48
on semiconductors and AI computing industry which
is drawing on a decade of my experience there are
00:12:55
7 modules in which I walk you through the
entire semiconductor value chain how it works in
00:13:02
details we discuss AI chips and alternatives to
NVIDIA GPUs will it ever be oversupply we will
00:13:12
also discuss exhorting technologies and its
potential like analog computing neuromorphic
00:13:17
computing probabilistic computing and what
the future holds beyond silicon if technology
00:13:24
and investing in technology is something you are
interested in check it out through the link below
00:13:31
the first 50 people to sign up with my code EARLY25 will get 25% off and if you register right now
00:13:41
you get a bonus a guide to AI silicon startups a
very good one see there of course when I hear 100
00:13:49
million times more energy efficient that's a lot
but from my discussion with Kerem I understood that
00:13:56
in order to build to emulate one probabilistic bit
on a classical computer with classical transistors
00:14:04
we need 10,000 transistors for a single bit
so when we want to get 1 million probabalistic
00:14:11
bits we need 10 billion transistors and then
there is also energy cost to generating random
00:14:18
numbers so emulating probabilistic bits running
probabilistic algorithms on a classical computer
00:14:25
is very energy intensive back to Extropic to the
most interesting part in the video you can see the
00:14:33
cryogenic lap that Extropic uses to fabricate
and test these devices what's strange at the first
00:14:40
glance this thermodynamic computer resembles a
lot a quantum computer and of course considering
00:14:48
the background and experience of the founders
totally makes sense if you want to know what
00:14:53
people are working on check out their experience
and when you see this footage the first question
00:14:59
which comes to mind is why to cool it down
if we are harnessing noise right in case of
00:15:06
a quantum computer the reason we have to keep it
cool because we want to get rid of the thermal
00:15:12
noise which appears at room temperature while in
a thermodynamic computer we harness noise so why
00:15:20
to cool it down why to make it so complicated if
theoretically it works at room temperature in this
00:15:29
case it has to do with super conductivity
which they harness through JJ's for computing
00:15:34
and which occurs only at low temperatures as we
discussed before there are many ways to build a
00:15:41
probabilistic bed using CMOS technology or magnetic
memory which is already very well scalable so
00:15:48
choosing here JJ's which rely on super conductivity
is a hard road to follow especially when when we
00:15:55
think ahead about scaling this system to 1 million
probabilistic bits in any case I love Extropic's
00:16:05
work I really admire them for pioneering a new
approach to computing because it's a very hard
00:16:12
job to be the first one you have to solve problems
no one thought about before like building a whole
00:16:19
software stack for this new hardware ground up we
discuss these aspects much more in details in my
00:16:27
course what makes me even more thrilled about this
thermodynamics technology that it seems there is
00:16:35
a way to build this JJ's Josephson junctions in
CMOS technology which means it's more practical
00:16:44
and more scalable it seems Extropic is also
looking into this direction and they've already
00:16:51
started working on their new thermo chip here
an estimation from Extropic of performance of
00:16:58
s such a silicon chip versus the performance
of a GPU and as you can see they estimate
00:17:05
orders of magnitude less energy consumption
per sample and time per sample what's very
00:17:12
important clearly probabilistic computers want
replace our traditional digital computers because
00:17:21
simply there are many applications where we need
absolute precision like take banking transactions
00:17:29
or heart pacemaker you don't want any uncertainty
here here we will still rely on traditional
00:17:37
digital computers however for applications where
probabilistic algorithms have a huge advantage
00:17:45
like in neural networks where the whole set of
problems have probabilistic nature or simulation
00:17:52
of natural processes for example Monte Carlo
simulations which is so popular on Wall Street
00:17:59
here I can see that probabilistic computers
will shine in any case it will take us
00:18:06
quite some time to build the hardware and the
software stack for it and then to actually in
00:18:14
practice demonstrate probabilistic supremacy
if you enjoyed this video share it with your
00:18:22
friends colleagues and on social media and if
you're interested in getting insights into
00:18:29
the semiconductor industry and investing in
silicon check out my new course and remember
00:18:35
to use the code EARLY25 to get 25% off and to
get the bonus looking forward to see you ciao
