Three Ways to Extend the Life of Moores Law

Transcriber: yoojae lee
Reviewer: David DeRuwe

My mom is 76, and I am her tech support.

The most frequent issue she runs into
with her iPad and smartphone is data.

She has accumulated so much data
that she’s running out of space.

Now, on the other end
of our generational spectrum,

here comes my teenage daughter,

who took 50 pictures of the exact
same pair of her Converse shoes

just to be able to pick the best one
to post on her Instagram.

Consequently, my family’s entire
iCloud storage was used up.

For me, this is not just a personal
problem, but also a work challenge.

You see, I work at Facebook
as an engineering director,

and it is my job to ensure
there is enough space,

which is what I call cloud
capacity, for every user.

Today, my company
has about 2.6 billion users.

Let’s imagine each of these users
uploaded one photo.

If I print these photos
into our four-by-sixes

and stack them side by side,

the prints will have enough

so that they could circle
our entire beautiful Earth

not just once, but a total of 10 times.

To support these users,
we not only need to store all the data,

but we also need to process them,
which means we need computers.

Lots of them.

A technology research company
called IDC made a prediction

that data stored in the cloud
would reach 100 zettabytes by 2025.

That is a 100 trillion billion bytes.

At this rate, simply adding more computers
will no longer be a sustainable solution.

You may wonder:

Data has been growing fast
like this for quite a while,

so why is it a big deal?

You’re right.

I was never worried because I had
a superpower in my pocket,

and it’s called Moore’s Law.

So what is Moore’s Law?

It is a prediction that says
computers get faster for the same cost.

In 1965, Gordon Moore,
the co-founder of Intel,

predicted that the number of transistors

on the silicon chip
would double every year.

The semiconductor industry
took his prediction as their North Star,

and they met the challenge.

Ten years later, Mr. Moore
revised his prediction

for the number of transistors
to double every two years.

The semiconductor industry
met that challenge again.

When I first came to the US for graduate
school in 1994 from Inner Mongolia,

my professor told us about Moore’s Law,

and he basically said there’s no need
to worry about how to keep up

with the growing data
because Moore’s Law was taking care of it.

This has been true
for the last 50 years till recently.

There is now a physical limit
of the silicon chip getting reached,

and therefore, we now start
to see the slowdown.

To give you a rough example,

on a silicon chip
the size of my fingernail,

there can be up to 50 billion transistors.

50 billion - that’s 10 times the number
of trees we have on this planet.

Recognizing that Moore’s law
may no longer help us as significantly,

the entire cloud infrastructure
industry worked together

looking for new methods
to bend down that demand curve.

So here you can see, this green line here,
indicating our computer supply,

and this blue line is the increasing
demand to store and to process the data.

As you can see, the supply
and the demand are mostly kept up,

but you also start to see a gap,
and this gap is widening.

This means my dear mom may have
to delete her precious videos and photos.

This means you and I may have
to pay more for our iCloud storage.

This also means more and more
data centers are going to be built,

consuming more power, cooling,
and space on the planet.

Luckily, my colleagues
and I do have a few methods here,

each of them bending down
the demand curve a little bit at a time.

And here I’d like to share
three of them with you:

The first method we use
to bend the demand curve

is by introducing hardware accelerators.

Hardware accelerators
are like specialists in the hospital.

While a general doctor can be very good
triaging and admitting a lot of patients,

she scales herself by referring specific
patients to the specialist.

In the data processing world,

these can be video encoding

or an intensive artificial
intelligence calculation.

Adding the hardware
accelerator to a computer

is like adding specialist support
to a general doctor,

and working together,
they become much more efficient

by focusing on what they’re best at.

To give you a rough example,

if I tell my computer to recognize
the image of my daughter’s shoe photo,

it would have taken about a day
using a regular CPU,

but with hardware accelerators,
this can be done under 30 minutes.

Hardware accelerators help
to bend the demand curve.

The second method to bend the demand curve

is by using software
to optimize our cloud platform.

Now, platform optimization
is like playing the game of Tetris.

The different colors of tiles

are like different types
of computers in my data center.

Some are for storage,
while the others are for compute.

I would spread and stack them

so that there’s as little
wasted space as possible,

which means I get to have less idle
computers and a higher utilization rate.

This Tetris game
is so fancy that it’s even 3D,

which means I get to leverage time zones.

When the users from the continent
of Asia went offline,

I can possibly use
those computers to process data

for our North American users.

And just like any gamers,

I would keep track of my gaming scores
from one day to the next,

and each day I get to bend
that demand curve a little bit more.

Now the third method, also the newest,
is to change how the data is used,

therefore, where the data is stored.

For example, there is a technology
called “on-device compute.”

Instead of uploading
all the data up to the cloud

and then process them up there,

we would process the data
locally on your devices,

either your iPhone or your laptop,

and from there, only a subset of the data

after processing
gets uploaded to the cloud.

If you imagine my data centers
literally as a box,

this approach is literally
thinking outside the box.

By reducing the amount of data
uploaded onto the cloud,

we get to bend that demand curve
a little bit more.

To sum it up, as you can see,
there are multiple technology inventions

being used to address
this steep growth curve of data.

As a result of applying
hardware accelerators,

software optimization,
and on-device compute,

all the cloud computers
are holding strong for all that usage.

If I can give you an example,

in the last eight minutes or so
since I started this talk,

more than 4000 hours worth of video
got uploaded to YouTube,

about 120 million messages
were sent by text,

and my daughter may have just uploaded
another 50 photos of hers.

Keeping up with all the data growth
will continue to be a challenge,

but I am optimistic
that we will continue to have

new inventions being
developed along the way.

I’m committed to support
all of my cloud users,

ensuring there is enough space
to carry their stories.

I’m committed to be my mom’s tech support,
guarding her photos and videos,

but I do plan to go talk to my daughter

about her thousands
of the same shoe photos.

Someday, even when I no longer
have the support from Moore’s Law,

I’m confident that my colleagues and I

will find ways to operate millions
of our cloud computers,

and operate them well,

because these computers
are indeed so important to us.

They carry not only all the data,
but also our memories,

our laughter,

and the stories of our lives.