On this episode, Omar Zarka, GM of AWS Panorama, dives deep into the structure of AWS Panorama, a machine learning appliance and software development kit that brings computer vision to on-premises internet protocol cameras. Omar discusses how AWS Panorama can make accurate predictions, how to reduce operational overhead, and how to improve the experience for customers.
AWS superfan and CTO of ESW Capital
GM of Amazon Panorama
Hi, everyone, welcome to another episode of AWS Insiders. We have a very exciting conversation lined up today,
about a relatively new edge computing service, called AWS Panorama. And my guest today is the GM of AWS
Panorama, Omar Zarka. Omar, it’s an absolute pleasure to have you on the show today, and I’m really excited to
learn more about the service.
Thank you, Rahul. I’m excited to be here. Thanks for having me.
Awesome. So, to get started, we’d love to know a little bit more about your Omar. Why don’t you tell us a little
bit about your history with AWS and how you started on the Panorama project?
I’ve actually been with Amazon now for nine years. In April, it’ll be nine years. I started out at lab 126,
which is the devices division. So the Echo, the tablets, the Kindle e-reader, Fire TV, et cetera. And I joined
when that team was relatively new and growing very, very fast, and got to see the all kinds of different product
ship, hit very important milestones, like millions or tens of millions of devices. And I learned a lot
specifically about devices, which became very important for my journey here on Panorama. And as I was reaching
about seven and a 1/2 years or so, I reconnected with a colleague that I had worked with at lab 126, who was
leading AWS Panorama at the time. And we connected, we were talking about Panorama, which I found it very
exciting, computer vision at the edge. And he thought that they had some roles that would be a good fit. So, one
thing led to another and I joined the team.
That’s awesome. So what was the Genesis of Panorama as a service? Was there a particular problem that you were
trying to solve or did the service have more organic roots?
That’s a great question. The vast majority of services at AWS, it starts with a customer signal. And so the
Panorama one was particularly interesting in my opinion. We didn’t start by trying to solve a business problem.
We actually started by trying to solve an education problem. So as AI and ML became more prevalent in the cloud
with deep learning revolutionizing how easy it was to develop models, there was a gap, a skillset gap. In what
practitioners and builders were able to do, and the skills they needed to be able to take advantage of these new
technologies. So, what AWS they did is they created the programs to educate people on these new concepts so that
they could get hands on and see how they could apply to their particular context for business problems. One of
the first initiatives was called Deep Lens. And Deep Lens was a fully packed edge camera in a box, with a edge
compute, that allowed you to run machine learning models.
And the idea was to give folks a platform to learn what it meant to run these deep learning models on computer
vision use cases. It was a revolutionary idea. People loved it, just like all kinds of interesting use cases,
whether it’s sign language interpretation, or homework help or reading a book for kids, very creative
applications. One of the other things that came out of it though, is businesses started taking Deep Lens,
businesses started buying them in large quantities and trying to deploy them into their business processes. So,
the team would see these orders and be like, “Well, we should reach out and figure out what they’re trying to
do, because the pattern doesn’t quite match.” And as they got deeper, they learned that customers were intrigued
by the simplicity of Deep Lens. The all in one package that allowed them to easily deploy machine learning or
computer vision to the edge for their business problems, was very compelling.
They dug a few layers deeper, and for those who know a little bit about Panorama, we actually ended up not
building a camera, we ended up building what’s called an appliance. And it’s like a computer. It’s like a
mini-server, that you can attach to your network, and it turns out that it’s very expensive to replace cameras,
or to change out the infrastructure. And so the idea of being able to take your existing cameras and connect
them to a smart appliance that makes it easy to deploy these rich applications that can do the visual inspection
that customers want was very compelling. And that’s how the journey started.
That sounds really interesting. I think I’d like to dive a little bit deeper and talk about edge computing. The
vast majority of AWS services are based on cloud site compute, right? And we’ve all come to see the benefits of
its costs, the scale, the elasticity, as well as security. But in the last two or three years, there’s been an
explosion of edge compute. And edge computing services on the AWS side. How does AWS think of edge computing, as
compared to the more traditional cloud-site computing, that we’ve all been familiar with?
So I’ll start by saying, when we talk to our customers, we actually see that they have several use cases at the
edge. And there’s no one size fits all I would say. And you actually see this in AWS’s portfolio, where you have
a variety of edge solutions that actually try to solve the edge or hybrid, or edge to cloud or cloud to edge
problem, different ways. Whether it’s the snow family of devices that are about getting data at the edge and
then getting them into the cloud, or you have outpost, which extends the cloud into a customer’s facility. And
there’s a lot of others like IOT and green grass and SageMaker Edge, et cetera. But in the case of Panorama, the
signal that we saw was a little bit different. We saw that customers had use cases that needed to bring the
power of AI in the cloud, that’s been going on for five or six years, but they needed it specifically at the
edge, with the scalability that the cloud brought.
And the reason for that, is there are just some use cases that are more suited for the edge. And so I see three
signals when I talk to customers. And we actually challenge. By the way, when I talk to a customer, I challenge
them. Why do you need the edge? Because the edge as we’ll probably get into later, is in many ways more
complicated, right? And that’s not a bad thing, but we want to make sure that customers are ending up with the
right solution to solve their specific business phone. But so the three signals that I tend to see are one, is
cost. And this is why specifically cameras are interesting, because cameras produce a lot of data, right?
And exorbitant amount of data. That’s actually very expensive to get to the cloud for a number of reasons. One
is just the bandwidth cost of streaming that data constantly. And that assumes you actually have the connection
necessary. Many facilities, manufacturing facilities, boats in the ocean, et cetera, do not even have the
option, right? Of getting that kind of connectivity. So, that’s one barrier. The second barrier is latency and
reliability. And what I mean by this is, some use cases require sub-100 millisecond latency, right? You need to
be able to respond to an event, such as a worker safety event, or a safety event, such as a machine potentially
colliding with something, or the like.
Or you may have a factory, that every minute on the factory or every 10 minutes or every hour is millions of
dollars of production. And being down for 30 minutes is just not acceptable, right? So that’s the second factor.
And then the third factor is actually data regulation, right? So having more control over where your data goes
and doesn’t go, the ability to discard data immediately after processing it, or storing it locally, instead of
send it to the cloud to have better controls, the third factor.
That sounds like a really good framework to decide whether you should go for edge computing, or just rely on the
elasticity and the cloud economics, basically. In general, do you find that edge computing is more expensive
than the cloud computing?
I would say that for all the reasons why the cloud has revolutionized the entire world over the last 20, 25
years, are the reasons why you should choose the cloud first, right? And basically, it’s simple. You don’t have
to buy hardware up front. It’s not a capital expense. It’s operational expense. Your ability to scale your
elasticity is super high. Your time to market and time to value is super fast. Your ability to control costs can
be done day over day, as opposed to you don’t have to plan them with hardware purchases over six, nine, 12
months. So, the cloud is infinitely more flexible, but the edge has some use cases that are just not solvable
right now in the cloud. There’s going to be a while where at least until our network infrastructure is strong
enough, where everywhere in the world, more or less, can have the bandwidth required to use the cloud. And,
that’s many, many, many years out there, there’s a lot of opportunity here for us at the edge.
Do you see that changing if solutions like 5G become more commonplace?
I think the dynamics will always change. But I think if you look at this historically, the swing between cloud
and edge and cloud and edge and distributed versus local has happened repeatedly over time. So, I think there’s
always going to be value and use cases at the edge, for a variety of reasons. And we want to meet customers
where they are, based on their needs, so that we can provide them those solutions.
Let’s start by talking about the hardware. So what is it that customers install on-site? And what is in that
Yeah, so we basically have a little box. It’s the size of a laptop, maybe a little bit thicker, but the same
dimensions. And the box connects to your network over the ethernet. And the basic premise is, you want to be
able to stream data into the box, so that you can process that data. And we are focused on computer vision. So
IP cameras, so a lot of businesses have cameras distributed throughout their facility for a variety of use
cases. So by connecting to the local network, you can tap into that video data, and then stream it through the
Panorama appliance. And the appliance actually doesn’t have very much storage. We have some other hardware
variants that will come out with more storage in the next month or two. But right now, we’re really focused on
processing. Because that’s the main use case. That’s the main gap.
A lot of customers already have storage on site, but the main gap that they have today is they can’t do this
intelligent processing. The main hardware component is Nvidia Jetson, Xavier AGX. Which is a system on a module,
produced by Nvidia, that is designed for AI and machine learning workloads. It leverages some of their most
popular GP architecture. And they’re very popular software stacks, which is deep stream, et cetera. And it
allows customers to deploy those models to this device, so that they can do things like detect objects, identify
when there’s a collision about to happen, identify when there’s a pattern that’s not expected, identify objects
that are out of specification, or need to be handled or inspected differently. A variety of use cases. Almost
anything that you can think of, that you can do with visual inspection, you can almost find a way to do it with
computer vision these days.
That sounds like a really, very tiny box that you have to install. And I hear it’s weatherproof and can be
installed outdoors and pretty resilient.
Yeah, we are dust and water resistant. So a lot of environments, unless you’re in a heavy water environment,
chances are, it’s going to be fine. Also for folks that are like, “I want to install this in a server room.” It
installs into server racks. You can put two side by side, it’s called one-U, single unit. One of our customers
has these 10 of them stacked right next to each other in their data center. It’s very interesting to see. But
yeah, it’s very versatile. We designed it with a lot of manufacturing and industrial use cases in mind.
What is the elasticity of these units? How many cameras or streams does each unit handle?
Our general guidance is to assume between 10 and 20, but we have customers getting more out of it. And it really
depends on the workload that you’re trying to deploy. In principle, the Nvidia chip that’s on here, can decode
dozens of streams at a time. What tends to be the bottleneck is the complexity of the machine learning model.
And that’s very use case specific. We’re working with customers to actually help them maximize the value of
their hardware, to be able to stream cameras intelligently or swap between different cameras. Our goal here is
to make sure that customers get the most out of their investment, and are able to scale a familiar way, as they
would in the cloud. Where they can add more cameras and more applications and distribute across multiple devices
So, I see there are multiple different elements. There’s of course the hardware part of it, which you install on
prem. Make sure that this device is connected to all your IP cameras, that you want to stream out of. But then
the second 1/2 of the equation is your machine learning models, and what you want to do with it. So, what does a
customer typically go through when they get started? How do they go about training it? How do they typically
capture the data from all of their cameras? What approaches do customers take?
There’s actually a few layers to that. The first is actually, one thing that customers love a lot about the
Panorama experience, is the ease to set up the device. Right now, if you want to take a device and connect it to
IOT, any web service, there’s actually a lot involved. And to do that in a scaled way, requires a lot of heavy
lifting, the customer needs to do. And to layer on top of that, to do it in a secure way is even harder. So,
we’ve abstracted all of that for our customers. When you buy a device, in five minutes you have your device set
up on your network. Not only is the device software stack secured top to bottom, but it’s connected to our
service, and you have a reliable connection to the service. We’ve taken days or weeks of work, I would say
sometimes even months, depending on the security profile that you’re trying to achieve, down to just a few
minutes for customers. So, that’s one really important thing.
So once you’ve provisioned your device, like you said. You connect it to the network, you’re able to identify
the cameras, but the main thing you want to do, it’s a platform to build these applications. And the core of the
application is the model, as you suggested. So, we can talk about maybe the model and the applications, and then
connecting the cameras. So from a model perspective, there are two big things that a customer needs to think
about. One is, you need a make sure that the model you’re selecting and you’re going to train is going to solve
the use case that you want to solve, right? And so there’s a lot of models out there, there’s a lot of
selection, and there are a lot of different ways to train your model to do that.
So we try to simplify this for customers, by doing two things. One is we pre-select models that we qualify on
our device. And we provide these in of sample applications. So these are open source models that are easy to
retrain for specific use case, and customers can deploy these also in a matter of minutes. The second thing we
do is we have really deep relationships with partners, who are experts at retraining and deploying these models.
It turns out that a lot of customers don’t have the data science expertise to be able to retrain these models in
a way that will solve their use cases in a satisfactory fashion. So, we work with our partners, to develop these
models or to retrain these models such that customers can then use them for their use cases. So, a typical flow
would go like this. Customer buys a device, they deploy their device in a lab environment where they can either
access a couple cameras or the like.
And they usually start with a sample application, just to get a sense of, “Hey, how does detection look like?
What kind of data am I getting back? Et cetera.” And then, we go through a process with the customer of
identifying, okay, for your use case, what does success look like? Right? Are we trying to achieve very high
accuracy? Are we trying to achieve very low latency? Are we trying to just have an integrated system, where you
used to have a disconnected environment, and now you want to get data back to the cloud? Maybe not necessarily
images, but inference data. And based on those factors and the number of cameras they want to connect in the
environment, we help identify the model, and then train the model for them, with our partners.
We don’t do a it as AWS, but the partners will do it in collaboration with the customer. And so then you end up
with a proof of concept. And that proof of concept allows customers to see the business value that they can
derive, from using machine learning at the edge, or AI at the edge. And then we start to talk about, okay, what
does it mean to actually deploy now? Because you need to do you need to do everything we did so far, but you
need to do it across tens or thousands of sites, right? So you need to be able to manage fleets. You need to be
able to manage models at scale. And so that’s the core of the prototype. And then once you get into building
something more complex, now we start to think about, okay, how do I attach cameras at scale? How do I package
and bundle and deploy my applications? How do I connect them with my other business systems? What kind of
dashboards do I want to be able to see and manage in the cloud, alerting, monitoring, et cetera?
So it tends to go, devices and provisioning, making sure it works on your network and works for your use case.
Then building the model that makes sense, and then applications and the dashboards, et cetera.
And if I understand right, you can have one device actually deploy multiple models?
So that on the same set of streams that are coming in, you could actually do multiple different kinds of
analysis, triggering different alerts. Is that accurate?
That’s absolutely correct. Yes. It’s like any computer, right? You can max out the CPU or the GP more or less,
but yes, we find that so far, most customers are able to run multiple models, and still serve their use case.
There’s always a trade off that you need to make as you would with any compute environment, but it very doable.
That’s why we picked the AGX, by the way. The AGX is one of the most powerful edge chips available on the
What are the different kinds of models or analysis that are commonplace? Is there a marketplace for standard
models that you could possibly apply right out the box, and start using them? And you see customers using them,
or has everyone go in to build custom models for their use cases?
So, I think the pattern that we see is that everybody starts with some sort of open source model. Either from a
model zoo that we have connected, or we provide guidance on how to pull those models into the Panorama
experience and deploy them to our device. Or with the sample applications. So, our sample applications represent
the most common use cases that we’re seeing so far. So that usually gets customers started. But what we notice
is that once you have a general proof of concept, to really get the business value that you want, you really
need to spend time training for your specific use case. We haven’t found yet, a model that is just a catchall,
where you can deploy the single model as is, and it just works. And we talk very closely with our solution
providers and et cetera, and this seems to be a common thing in the industry in general. That there needs to be
some amount of retraining or training specific to the use case that the customer has.
What are your favorite use cases that wouldn’t really be possible if Panorama wasn’t around?
So my favorite use case I can’t talk about yet, unfortunately. But I plant that seed because I hope to be able
to talk about it with you in the next few months. But there’s a lot of really interesting stuff that you would
never expect, that make the lives of everybody easier and for the good of everybody. So, I hope to be able to
talk about more of them, but one of them, one of the unexpected use cases that I can talk about today is AI for
animals that Deloitte in the Netherlands created. And the idea here is to detect animal abuse in real-time. And
so they want to be able, it’s not just about identifying when the abuse has happened, but it’s also about
intervening to minimize the abuse. So, to me, this combines a lot of the advantages of edge AI and it is for the
general good of the world, right?
So, one is, it’s very hard to have humans watching all the time. Two is, humans themselves are part of the
problem at times. And then three is being able to respond in real-time or in locations that don’t have the
bandwidth or the infrastructure necessary to be streaming videos to the cloud, make it a perfect storm.
Can you give some examples of industrial applications of Panorama? I assume that those are far more common, and
so what kind of scenarios are people using Panorama in?
So the first thing I’ll say, and this might resonate with some of your listeners, is I think we’ve all felt the
pinch of the supply chain crunch, right? So, right now, supply in almost every industry is a problem. And
actually, a lot of that challenge is sometimes not the supply itself, but the ability to get goods to where they
need to go, in an efficient way. One of the use cases that we’re consistently seeing across industries that it’s
resonating with customers, and that customers are deploying into production, is the ability to optimize the use
cases. Using Panorama to get visibility as to where containers are in the process. So that they can in real time
inform their customers about where their containers are, how long they’ll take the process, et cetera. But also
gives them the insight they need to see where they have bottlenecks in the process.
Similarly, completely different use case, completely different industry, but similar use case, Tyson, in their
production of foods, need better insight as to which skews, which products are being produced at which rate? And
where they’re seeing slowdowns in their factories. So they use computer vision, to be able to do this in
real-time, so that their operators can make more intelligent decisions about where to spend their time across
which lines and which areas, et cetera. We also see again, similar use case, but very different industry.
Cincinnati Airport is using Panorama computer vision, to be able to monitor curbside at airports. I think a lot
of people have the experience of trying to drop somebody off or pick somebody up and you just can’t get to the
curb, right? And you have to loop back around.
And so the idea is very simple. Instead of having somebody standing there, intimidating everybody and just… I
can only imagine what it’s like to be a person monitoring the curbside. So they’ve abstracted all that by just a
very simple solution. They just see cars coming in, and they just time. They just set a timer. How long has that
car been there? And if a car has been there, passing whatever threshold, they send a notification to an
employee’s Apple Watch, and they go politely tell the car to keep moving. It doesn’t have to be binary of
replacing something, or there’s one benefit and then you lose something else. So yeah, just between those four
use cases that we described, you can see the core is very similar. It’s like, there’s an object, I’m tracking
something or I’m tracking some behavior, but then the use cases and the consequences to the business are very
Do you see applications of Panorama in areas like healthcare and sports as well? But that seems to be a very
interesting domain. You started seeing AWS themselves are involved in a lot of sports broadcasting now, where
there’s streams of data coming in and there’s tons of AI being applied right there on everything from statistics
to probabilities of certain outcomes as the game is being executed. Is Panorama being used in scenarios like
that. Or are you seeing applications like that evolve?
So, we are seeing applications like that evolve. We have been in conversation with multiple customers in
healthcare and in sports, trying to identify how we can make their solutions better for their customers. And it
could be anything from, in the healthcare space for example, fall detection, right? Identifying when somebody
needs help or on sports. There’s a ton of applications coming out now. And AWS is pioneering with the NFL in
this space. There’s nothing yet, but I would say there’s a lot of opportunity here and we’re excited to work
with our partners and customers there.
I want to jump in to ask you about the three best practices that you’d recommend customers follow when they
decide that they want to start with AWS Panorama. What will be the three pieces of advice that you’d give them,
to get started?
Yeah. Great question. So the first advice that I give to a customer is to engage with a practitioner, that
understands how CV, computer vision, works at the edge and what can be enabled. And so the reason I say that is
there’s a lot of brainstorming that happens around the use case. All of these use cases that we discuss, the
core is the same, right? The core is very similar across all of them, but the actual implementation, how you get
it done and what are the important differentiators for that use case in that particular situation, can vary a
lot. So I think this is where our practitioners and our partners really help change the conversation, and
identify the best way to achieve a particular solution. So, that’s the first thing I would say. The second thing
I would say, and this goes back to maybe earlier in our conversation is, really think about why the edge, right?
There’s a lot of benefits to the cloud. It’s funny because, my business is for the edge, but at the same time,
we really push our customers on why the edge? Because if you want to go down this route, I think your businesses
are seeing dramatic improvements in their business processes and what they’re achieving, but it has to fit,
right? You don’t want to fit a square peg into a round hole, et cetera. So really be precise about why the edge
matters in your use case. And we’re happy to help you figure that out. And as well, the partner.
I think the framework that you described earlier is a really good way to think about why edge, if at all.
Yes. A 100%. No, for sure. Just to recap, right? Cost, usually driven by bandwidth or lack of availability of
network. The second is latency as in needs to be real-time, or you need to have very, very high reliability for
processes that require one millisecond processing times or hundreds of milliseconds of processing times. And
then the last one is really refined control of data, right? On where data ends up and how you use data, those
are the three that really make a big difference.
Great. And was there a third best practice recommendation you had?
I mean the third one, it’s like the first, but I think it’s important for folks to realize is that, a lot of our
customers who even have a lot of data science expertise, are able to get through the initial phases, technical
validation, business validation, on their own, which is fantastic. But really to scale, if you think of the edge
at scale, when you think of AWS and the way you want to scale, provisioning, it’s so easy, right? You just add
instances, there’s very little thought that you have to put into it. Here at the edge, it’s very different.
Managing the fleets at the edge and rolling out the networks, et cetera. So it can be complicated and can
require a lot of boots on the ground, in a lot of geographies that not all customers have.
So because of that, I encourage all of our customers to work with partners. That’s where we see the best
success. And we have a great partner network of partners who have a lot of experience in the space, across a ton
of use cases that can help customers be successful.
I think you’ve covered some of it, in terms of some things to consider and not do. But if you could specifically
dive deep into what are the cost related practices that you would advise our customers to follow as best
practices for the service?
So, I think in our case with the edge, starting by doing the technical and business validation before you scale,
it seems like an obvious thing, but I’m going to point out why. The reason why it’s more important is because
there’s a hardware purchase involved. Actually doing the profiling upfront, helps a lot in being able to
understand the devices you need and how you need them. Whereas in the cloud, you have a little more fungibility
in terms of if you underestimate or overestimate, you can deprovision or reprovision. So, you want to be a
little more intentional up front on the hardware. But, for what it’s worth, I think we haven’t seen yet a
customer run into this problem. I think that the natural cycle of how this goes at the edge, ends up where most
customers are able to get ahead of this. We do see some customers that want to jump ahead, they’re excited,
which is fantastic.
We try to make sure that they understand the importance of doing this up front and we walk them through. I think
that is the main one. I mean the rest is pretty straightforward. We only bill by camera, right? On the number of
cameras that you have connected. So, if you’re getting value out of cameras, have them connected. If not, it’s
very easy to disconnect them and we stop billing for them. We were very, very intentional about keeping the
billing very simple, because of exactly what you described. And I think by keeping it just along the dimension
of charging per camera, has helped mitigate some of these cost overruns that could potentially happen.
Yeah. I think what we see across a large number of other AWS services is that the elastic nature of the service
itself, where you auto scale up or down, causes that build shock once in a while, because you don’t expect the
system to scale up suddenly, it is unanticipated and it probably didn’t scale down quite as fast as you
anticipated. And that causes large build else to accumulate. But in the model that you guys have now for
Panorama, it looks like the costs are actually very predictable, because it all depends on the number of cameras
that you add, or the streams that you add for analysis. And it’s just that. And that’s not something that you
see as being elastic in nature.
You don’t auto scale it up the number of cameras or stream. So, that seems like a pretty good fail safe that’s
built into the service pricing itself. Great. And one last thing. So how does someone start with the hardware? I
hear you can buy this now on the Amazon store, but what are the mechanisms of getting hold of the hardware?
We made it very simple. You can literally buy on amazon.com with your amazon.com account. You can buy it with
your personal. So, for those who don’t know, on amazon.com, there’s two ways you can have your personal account,
it’ll be shipped to you in two days if you have prime. Or you can use Amazon for business, which allows
businesses to use the purchase order process, so that they can match their business processes. We also have
what’s called AWS Elemental, which is a manual purchase order workflow that’s supported through the AWS console,
for those that prefer that also.
And the hardware is going to be Amazon only, or are they plans to have other partners offer the hardware as
We’re very much keen on working with our hardware partners, to enable more skews and more variety for our
customers. So, Lenovo is the first to launch. They’re launching here very soon. We announced with them a couple
times last year, they are producing a variant of this device, different form factor, but also different chip. So
it’s the same Jetson family of chips from Nvidia, but it’s what’s called BNX, which is a lower price point, but
still very powerful and very applicable to a lot of use cases.
I have to say that this has been an absolutely fascinating conversation. And thank you, Omar, for being so
generous with your time, and sharing your insights with us. I would love to have you back on the show, whenever
you’re ready to share with us and discuss the top secret and unexpected use case you spoke about earlier for the
audience, I encourage you to go and try out AWS Panorama. And we’d love to hear from you about the use cases
that excite you. If you enjoyed this conversation, please take a few moments to review the show. And I hope to
see you on the next episode of AWS Insiders. Until then, goodbye.