covering AWS re:Invent 2017.

Presented by AWS, Intel,

and our ecosystem of partners.

>> Hey, welcome back everyone.

We are here live at AWS re:Invent in Las Vegas.

This is 45,000 people are here inside

the Sands Convention Center at the Venetian,

the Palazzo, and theCUBE is here

>> Offscreen: I don't have an earpiece, by the way.

>> for the fifth straight year,

and we're excited to be here,

and I wanna say it's our fifth year, we've got two sets,

and I wanna thank Intel for their sponsorship,

and of course our next guest is from Intel.

Scott Macepole, director

of the CTO's office at Intel PSG.

Welcome to theCUBE. >> Thank you.

>> Thanks for coming on.

So, had a lot of Intel guests on,

lot of great guests from customers of Amazon,

Amazon executives, Amy Jessup coming on tomorrow.

The big story is all this acceleration.

of software development.

>> Scott: Right.

>> You guys at the FPGA within intel

are doing acceleration at a whole nother level.

'Cause these clouds have data centers,

they have to power the machines

even though it's going serverless.

What's going on with FPGAs,

and how does that relate to the cloud world?

>> Well, FPGAs I think have a unique place in the cloud.

They're used in a number of different areas,

and I think the great thing about them

is they're inherently parallel.

So you know, they're programmable hardware,

so instead of something like a GPU

or a purpose-built accelerator,

you can make them do a whole bunch of different things,

so they can do computer acceleration,

they can do network acceleration,

and they can do those at the same time.

They can also do things like machine learning,

and there's structures built inside of them

that really help them achieve all of those tasks.

>> Why is it gonna pick up lately?

Because what are they doing differently now with FPGAs

than they were before?

Because there's more talk of that now more than ever.

>> You know, I mean, I think it's just finally

come to a confluence where the programmability

is finally really needed.

It's very difficult to actually

create customized chips for specific markets,

and it takes a long time to actually go do that.

So by the time you actually create this chip,

you may have not had the right solution.

FPGAs are unique in that they're programmable,

and you can actually create the solution on the fly,

and if the solution's not correct

you can go and you can actually change that,

and they're actually pretty performant now.

So the performance has continued to increase

generation to generation,

and I think that's really what sets them apart.

>> So what's the relationship with Amazon?

Because now I'm kinda connecting the dots in my head.

Amazon's running full speed ahead.

>> Scott: Yeah.

And they're moving fast, I mean thousands of services.

Does FPGAs give you guys faster time to market

when they do joint designs with Intel?

And how does your relationship with Amazon

connect on all this?

>> Absolutely, we have a number of relationships with Amazon,

clearly the Xeon processors being one of them.

The FPGAs are something that we continue to

try to work with them on,

but we're also in a number of their other applications,

such as Alexa, so and there's actually technologies

within Alexa that we could take

and implement either in Xeon CPUs or actually in FPGAs

to further accelerate those,

so a lot of the speech processing,

a lot of the AI that's behind that,

and that's something that, it's not very prevalent now,

but I think it'll be in the future.

>> So, all that speech stuff matters for you guys, right?

That helps you guys, the speech,

all the voice stuff that's happening,

and the Alexa news, machine learning.

>> Right.

>> That's good for you, right? I mean, that, I mean...

>> It's very good, and it's actually,

it's really in the FPGA sweet spot.

There's a lot of structures within the FPGAs

that make them a lot better for AI than a GPU.

So for instance, they have a lot of memory

on the inside of the device,

and you can actually do the compute and the memory

right next to where it needs to be,

and that's actually very important,

because you want the latency to be very low

so that you can process these things very quickly.

And there's just a phenomenal amount of bandwidth

inside of an FPGA today.

There's over 60 terabytes a second of bandwidth

in our mid-range Stratix 10 device.

And when you couple that together

with the unique math capabilities,

you can really build exactly what you want.

So when you look at GPUs, they're kinda limited

to double precision floating pointers,

single precision, or integer.

The FPGAs can do all of those and more,

and you can actually custom build

your mathematical path to what you need,

save power, be more efficient,

and lower the latency.

So...

>> So Andy Jessup talked about

this is a builder's conference.

The developers, giving the tools to the developers

they need to create amazing things.

One of the big announcements was

the bare metal servers from AWS.

>> Scott: Yeah.

How do you see something like an FPGA

playing in a service like that?

>> Well, the FPGAs could use to help

provide security for that.

They could obviously be used

to help do some of the network processing as well.

In addition, they could be used

in a lot of classical modes that they could be used in,

whether it's like an attached solution

for pure acceleration.

So just because it's bare metal

doesn't mean it can't be bare metal

with FPGA to do acceleration.

>> And then, let's talk about some of the...

You guys, FPGAs is pretty big in the networking space.

>> Scott: Yeah.

>> Let's talk about some of the surrounding

Intel technologies around FPGAs.

How are you guys enabling your partners,

network partners, to take advantage of X86, Xeon, FPGAs,

and accelerating networking services

inside of a solution like Amazon.

>> We have a number of solutions that we're developing,

both with partners and ourselves,

to attach to our nix, and other folks' nix,

to help accelerate those.

We've also released what's called the acceleration stack,

and what that's about is really

just kinda lowering the barrier of entry for FPGAs,

and it has actually a driver solution

that goes with it as well, it's called OPAE,

and what that driver solution does,

it actually creates kind of a containerized environment

with an open source software driver

so that it just really helps remove the barrier of,

you know, you have this FPGA next to a CPU.

How do I talk to it?

How can we connect to it with our software?

And so we're trying to make all of this a lot simpler,

and then we're making it all open

so that everybody can contribute

and that the market can grow faster.

>> Yeah, and let's talk about ecosystem around data,

the telemetry data coming off of systems.

A lot of developers want as much telemetry data,

even from AWS, as possible.

>> Scott: Yeah.

>> Are you guys looking to expose

any of that to developers?

>> It's always something under consideration,

and one of the things that FPGAs are really good at

is that you can kinda put them towards the edge

so that they can actually process the data

so that you don't have to dump

the full stream of data that gets generated

down off to some other processing vehicle, right?

So you can actually do a ton of the processing

and then send limited packets off of that.

>> So we looked at the camera today,

super small device doing some really amazing things,

how does FPGAs playing a role in that, the IOT?

>> They do a lot of, FPGAs are great for image processing.

They can do that actually much quicker

than most other things.

When you start listening,

or reading a little bit about AI,

you'll see that a lot of times

when you're processing images,

you'll have to take a whole batch of them

for GPUs to be efficient.

FPGAs can operate down at a batch size of one,

so they can respond very quickly.

They can work on individual images,

and again, they can actually do it

not just efficiently in terms of the,

kinda the amount of hardware that you implement,

but efficiently in the power that's required to go do that.

>> So when we look at advanced IOT use cases,

what are some of the things that end-user customers

will be able to do potentially with FPGAs out to the edge,

of course less data, less power

needed to go back to the cloud, but practically,

what are some of the business outcomes

from using FPGAs out at the edge?

>> You know, there's a number of different applications,

you know, for the edge.

If you go back to the Alexa,

there's a lot of processing smarts

that actually go on there.

This is an example where the FPGA

could actually be used right next to the Xeons

to further accelerate some of the speech,

and that's stuff that we're looking at now.

>> What's the number one use case you're seeing that people,

what's the number one use case that you're seeing

that people could relate to?

Is it Alexa? Is it the video--

>> For the edge, or?

>> Host: For FPGAs, the value of accelerating.

>> For FPGAs, I mean, while there's usage

well beyond data center, you know.

There's a classic what we would call wire line

where it's used in everything today.

You know, if you're making a cellphone call,

it likely goes through an FPGA at some point.

In terms of data center, I think

where it's really being used today,

there's been a couple of very public announcements.

Obviously in network processing

in some of the top cloud providers, as well as AI.

So, you know, and I think a lot of people

were surprised by some of those announcements,

but as people look into them a little further,

I think they'll see that there's a lot of merit to that.

>> The devices get smaller and faster

and just the deep lens device has got a graphics engine

that would've been on a mainframe a few years ago.

I mean, it's huge software power.

>> Yeah.

>> You guys accelerate that, right?

I mean I'm looking, is that a direction?

What is the future direction for you guys?

What's the future look like for FPGAs?

>> It's fully programmable, so, you know,