Smarter Devices at the Edge: A Conversation with FreeRTOS Founder Richard Barry

Highlights:

[00:05:00] – What is FreeRTOS and How It Evolved at AWS

[00:09:00] – FreeRTOS on Cortex-A: Real-Time Performance at Higher Compute

[00:13:00] – Inside the Cortex-A320: Designed for AI at the Edge

[00:17:00] – Greengrass to the Edge: A Continuum of AWS IoT Tools

[00:25:00] – Security by Default: Pointer Authentication and Arm v9

[00:34:00] – The Future: Multimodal ML on Low-Power, Context-Aware Devices

Transcript:

Brian: [00:00:00] Hello, and welcome to another episode of the Arm Viewpoints podcast, where we explore topics at the intersection of AI and human imagination. I’m Brian Fuller, your host, and today we’re diving into the expanding role of real time operating systems in Edge ai. In a world that’s moving faster and smarter than ever from flashing LEDs to running machine learning models, the microcontroller has come a long way and no one knows that evolution better than today’s guest, Richard Berry, creator of free r tos and head of IOT Technology Strategy at AWS.

Joining me as well is John Thompson, who leads software engineering for arm’s IO OT line of business. Together we explore how free RTAs. Arms, new cortex. A three 20 Core and AWS Greengrass are redefining what’s possible at the intelligent edge. Let’s [00:01:00] jump right in, gentlemen. Welcome. Welcome. Before we get started, it.

It occurred to me that some of us, probably not John, have been in this industry since the days when you could get a microcontroller to flash an LED was a big deal, and now they’re handling machine learning models, running anomaly detection and doing all sorts of amazing things. So the March of Progress is astonishing and the super structure around this is nothing short of mind blowing.

I think now you’re gonna tell us about that. Throughout this conversation, but first, tell us about yourselves, and let’s start with our guest of honor. Richard, take it away.

So if you want to make me feel old, my first microcontroller project, I think 256 bytes of Ram. So there we go. You

Brian: don’t look that old.

So I, I sound like one of the people I used to talk to that [00:02:00] told me about the days of punch cards. Anyway, I’m best known as the original author of Free Artos, which I’ll tell you a little bit about today. And been working on that since around 2002. Not exactly sure when the first kind of mark in history is 2004, when it was registered on SourceForge.

This predates skit, of course it was in SVN in 16. I joined Amazon Web Services, internet of Things, A-W-S-I-O-T. And I’m currently working as the head of IOT technology strategy amongst other things, and freear still working on free. It’s part of the AWS Edge device, edge software family, I should say.

Got it. John, over to you. Thanks, Brian. John Thompson. I run the software organization at ARM within our iot line of business and as you alluded to in the intro there, iot these days, at least within arm, has a strange definition [00:03:00] because what we’re talking about today is a lot of capability that certainly I don’t think I necessarily think aligns with.

People’s traditional view of the iot market. We’re talking about a lot of machine learning capability and a lot of really incredible things that we’re bringing down to the edge. My part of the organization, we look after the software that we deliver alongside RIP that we bring to market. So we work a lot on.

The reference software applications that we provide with our IP and with our subsystems, and also look a lot at developer experience. So what kind of tooling do we need to be able to provide with the products that we bring to market to make sure that. We’re not looking at the architectural capabilities of the products that we’re delivering, but also making sure that those capabilities are really accessible by everyone who’s consuming them along the value chain from device vendors that may be [00:04:00] deploying that ip, but also to developers as well that are interacting with that capability and re really looking to squeeze more performance out of the devices that they’re deploying.

Do you ever sleep or is that not an option? I’ve been known to sleep. It has happened maybe less and less these days, only because we’re working on a lot of very cool things. But I’m very fortunate. I have a great team of people around me, and I think at all more broadly, we’re surrounded by fantastic people.

So if I don’t sleep, it’s for good reasons. It’s because it’s a very stimulating place to be. Yes, indeed. Yes, indeed. I think we’re both motivated by trying to abstract away some of the complexity. Get people to get customers applications to market as quickly as possible, I think.

Brian: Speaking of complexity, this particular episode is an easy one for me because you guys get to talk amongst yourselves ’cause you’re the experts.

So I’m gonna hand it off to John and I’m gonna listen eagerly.

Fantastic. I think first of all, thank you, Richard [00:05:00] for the time to join us. People who are listening that may not be familiar with the background of Free atos, can you tell us a little bit about what Free atos is and how it’s the role of Free atos within AWS and the role it plays as part of a WS services?

I like to think it’s, the name is self descriptive here, but the free part is easy. MIT licensed open source, so it’s very liberal licensed, again, trying to make things. Easy for people to use it. MIT licenses generally on, on company’s approved lists already. The side of it or real time operating system side of it is more up for debate, I would say.

It’s designed specifically to be very small and simple. It’s delivered as c source code to use it. You. Build the source code. Hopefully make that simple as well’s, no dependencies on anything other than the C library, and you statically link it into your [00:06:00] application. So it provides a realtime scheduler.

Into thread or tasks as we say, into task communication, primitives, that kind of thing. We say tasks rather than threads just to avoid confusion. There’s no concept of virtual memory. We st we stay away from that because, there are lots of systems Linux for what, which are serving that market very well.

But if you think of an operating system as being separate, the operating system being separate from your application. In the real time space, you can think of systems like QNX or VxWorks. It’s not like that. So many people would just refer to it or as a kernel or scheduler. How does it relate to AWS?

For many years it was delivered or provided just as as a kernel and as you imagine with.

Solving its user’s biggest pain points. It’s very much kind of driven by the user base. So we started integrating other [00:07:00] libraries. In particular UDP and T-C-P-U-D-P kind of evolved into TCP. That more recently is evolved into a dual IPV four six stack and multiple interfaces. It started in that kind of connectivity direction, and I should say even the TCP stacks designed to be very small and hopefully simple.

Became very popular. Many devices running free RTOs, many devices connecting AWS have an IOT service. So there was a kind of symbiotic relationship, if you like, in US coming together. AWS iot are obviously motivated to make sure that devices can connect to their service quickly, but very importantly, securely as well.

And so AWS bring. Investments in doing things like memory, safety proofs and this thing security certifications. More recently, actually a safety certification, which is all the work’s being completed for, we’re just waiting for the feedback from the [00:08:00] assessor. And then on the other side, the Fritos community gets a deliberate kind of distribution model where we try and make sure that everything in the free bucket is completely independent.

There’s nothing. There’s nothing specific to AWS about TCP or TLS, and they get the benefit from those investments that Amazon make without having to be an Amazon customer. Like I say, you can use free autos for any application you like, connect to any cloud service you like. Obviously we would prefer you connected to AWS I.

RTOSs in general, are something that people more often associate with microcontrollers. What are the kind of use cases that you see where people are looking to deploy free RTOs, perhaps on a cortex, A core as opposed to on a microcontroller as we may have seen in the past? Yeah this comes to what we were saying in the intro there about, or implying at least the increasing complexity.

So was the kind of sweet spot, if you like, was the. [00:09:00] Microcontroller because they are big enough to run an application of a complexity, which actually benefits from multithreading. Free will run on eight bits as well. Often you think, why would you you’re probably adding complexity rather than removing it, following what is requested by the community.

Some years back, low end cortex, a microcontrollers as well. And as well, I still think 32 bit. Is the majority of applications, but where you have applications where you need real-time behavior, high throughput, more and more of ML applications. So if we are looking at industrial automation, machine control, some cameras that have ML built in.

Automotive applications, there’s as applications require more compute and particularly more data and the capabilities open up these broader use cases. That’s really where we’re [00:10:00] seeing the cortex. A use with free atos as well, often in combination with Linux in Multicore soc and in that example, what’s the relationship between.

The art OS stack and the rituals stack, it tends to be use cases where you’re seeing the art os behaving in a way where you’re requiring isolation, or do you see a level of interaction there on the system where you may be looking for. Behaviors or certain actions on a system which then prompt you to interact with the cause that are running a richos?

Yeah, so Linux is obviously a very comprehensive system. Lots of people are familiar with it, and if, I always like to think that most applications, the majority. The majority of the requirements are not real time, depending on your definition. Of course, that’s our whole other debate. But when you boil things down, even realtime applications the hard realtime requirements are normally minimal.

So that, that kind of split of benefiting from all of the infrastructure and the knowledge that Linux [00:11:00]brings, but still be able to use. An SOC to isolate those requirements, which are truly real time. That’s, I would say the majority of use cases. You can get the graphics and the networking and all that from Linux.

Fine. There are other, I would say edge cases as well. One is the functionality into on a call where it’s much easier to reason about it. You need to prove the schedule ability, which is a word I always struggle to say, or if you need to isolate code, which perhaps needs a higher assurance level. So that’s another use case.

And also power management as well, where you can leave a realtime algorithm running and just wake up the more capable cause, when they’re needed. So it’s quite broad use cases. That’s great. That’s really interesting on the topic then of looking at iot use cases and iot use cases becoming more [00:12:00] capable to bring in the a three 20.

Which is you can give me the correct terminology, which is the first I believe the first arm Ben nine architecture, which is specifically targeting iot use cases and how that kind of relates to what we are talking about here with real time, in particular, some of the security features which are now default on that core and how that relates to I ot especially where we have.

Use cases with people have physical access to the devices. Yeah it’s really interesting. So Cortex A three 20 is our first V nine Ultra efficiency core. And so it’s been designed and optimized specifically to cater for edge AI use cases and IOT applications and markets that we see. And so from a performance perspective, we’re seeing a really significant jump in machine learning performance.

If we compare the A three 20 to what [00:13:00] was our previous ultra efficiency called, the A 35. In some cases we’re seeing up to a 10 increase in machine learning performance. Similarly, if we’re looking at scale of vector performance we’re seeing roughly a 30% higher performance also against the A 35, both in terms of performance, but also in terms of energy efficiency.

We’re seeing a lot of gains. If you compare the A three 20 against the A five 20, for example, which was, which is our high efficiency V nine two cpu. In some cases, we’re seeing 50% more energy efficiency as well. The reason that we’re able to achieve those uplifts. Mainly due to the V nine architectural features that were shipping in the A three 20.

So it’s a six to four bit architecture, and it’ll be really interesting to understand some of the use cases that you see and how that pertains both to free s and some of the other services the A WS provides. But to take a few examples, it provides native support [00:14:00] for virtualization so you can deploy hypervisors.

For industrial applications, for example. So that case that you just mentioned, where you may want to have the RTA stack running on one core and then perhaps a Linux stack running on the other cores, either for power management or for different system configurations. It provides a huge amount of flexibility.

So positioning the A three 20 as a product where we can. Deploy rich operating systems stacks such as Linux or Android, but also stacks such as free s is really interesting. You mentioned you mentioned security capabilities as well, and we can touch on that a little bit later on. The markets that we’re interested in within our parts of the business, within arm.

We get a lot of demand for the energy efficiency features that I just mentioned, but also security capabilities that are aligned with both industry standards and use [00:15:00] cases that we see in areas such as industrial automation, potentially some robotics use cases as well. And there’s a sweet spot there where you need really high compute capability, but you still want to have determinism and you still need to have a comprehensive.

Security capability that you can bring to market as well. So it probably also lines up nicely with the kinds of requirements that you might have from AWS. So when you’re looking things such as device fleet management or integration with a WS cloud, it might actually be segue actually for you to tell us a bit about some more services within aws.

I know you’re responsible for. Green grass as well within AWS and it’d be really interesting to understand the kind of things that you’re looking at with AWS Greengrass and how that might be relevant to some of the markets that we’re targeting with the Yeah, it’s interesting where we have a device which is more compute power and, lower power budget and at first glance you say [00:16:00] existing applications.

And run lower lower power cost, especially if they can do their work much faster. But of course, the reality is as we all know that more likely, those extra capabilities will all be used up by running more powerful applications, especially in the. AI world, it’d be interesting to see, just again the step changes in, the experience we can bring to edge devices using the extra capabilities there.

I’ve talked about free atos, but Freeto is. Part of a continuum of products, which actually starts with the libraries I mentioned earlier. Everything other than our TCP IP stack is independent of the operating system on which is running. In fact, they’ll, in fact, they’ll run without an operating system, even just bare metal.

And then we have various products which use the additional capabilities. The [00:17:00] devices they target to really raise the abstraction level. So when we talk about something like free adding in that multi-threading, then we have some products which are encapsulating everything that’s required for a secure connectivity, including all the provisioning workflows.

How do you get the credentials onto devices to make sure they can authenticate So. We have a product called express Link, which actually puts all that onto radio chips with a very simple interface. Then what you’re alluding to after that is Greengrass. So Green Grass is, allows you to do. Several different things to make your life simpler, including lifecycle management, overthe, air updates.

If I was to summarize it as succinctly as I can, I would say it makes your operating system kind of cloud native in a way. So it provides the connectivity of the authentication for you, but it goes, it allows you like the lifecycle management, not just of applications, but also of. [00:18:00] Models, and this is just part of it.

’cause it also allows you to run Docker containers or you can run native applications in green grass and make use of the communication services it provides for you. So if I carry on past Express link in our continuum of products where you have a new implementation of the Nucleus for which is written in.

C plus. And I say that’s a new implementation because the full Greengrass product is written in Java. So now you can see that we have series of products of a continuum that build on top of a common code base up to Greengrass light, a Greengrass nucleus light. And then you can take that step change if you are running on a maybe a gateway or something, which is much more capable.

In fact, some of our industrial Edge devices are running green grass as well, green grass, B two with the Java product. So it’s a kind of horizontal [00:19:00] product. It’s. Giving you any vertical specific functionality, but the intention is to make it much easier, do a lot of the undifferentiating different d undifferentiating work, some of the more nuanced pieces that just allows users to guess on and just focus on their primary core value.

And what kind of things do you need to be exposed to in terms of device capability? Are there any standard capabilities on a device that need to met in order for. Be running green, either the Java version of. Green grass nucleus light. Yeah, there there’s quite a big jump in, in the system requirements here.

Glad you brought that up actually, because I should have mentioned it. Thanks for that. So the Java version, you need about 256 megabytes of disks, both 96 megabytes of ram. So that’s, this is why I was saying, gateway devices, that kind of thing, and more, more powerful cameras.[00:20:00]

Really fits into things like.

Low cost cameras. So rather than the gateways we’re looking at embedding into devices themselves. It requires about five megabytes of ram and about five megabytes of disc. So you can see it’s a big jump. It’s much, much bigger than FreeRTOS, which is you looking at, a hundred, 200 kilobytes.

Up Greengrass nucleus light five megabytes, and then there’s a big jump up to Greengrass V two at 96 megabytes. But I think that fits the kind of product families that we see quite well. And do you see a similar sense of use cases across those three, or do you have a pretty distinct separation between the use cases that you’re seeing on free RTOs compared to nucleus light and compared to green grass And yes and no even.

Even Greengrass V two is built into end devices. I keep talking about gateways, but one of our biggest customers there is actually in an end [00:21:00] device rather than a gateway. I would say there was more of a split in the personas of the developers that are using them. Lots of low level device developers come more from a kind of electrical engineering background, and they are more likely to use something like Freeto for the, if you look at the same use case, more likely to use something like Freeto more familiar with building C code, configuring C code green.

The personas that would use green grass are more people that have. Grown up with applications, programming, Linux proficiency, and are more familiar with the application level than they are perhaps at the hardware interfacing level, if that makes sense. That’s really interesting. Does that change then, how you expose that capability to the developers that are consuming those services?

Or do you present that in a consistent way, the way you would go about building a free autos application? Because we have. [00:22:00] To make sure that it is not dependent on any particular build system, any particular compiler. It has to be a C compiler, don’t get me wrong, but does it doesn’t matter the source of the C compiler.

You can just go to GitHub if you are familiar with that and clone the libraries. Or you can use something like simis packs or you can use even like package managers. Partners deliver it in lots of different ways, but they all assume that you are happy to deal at that low level to deploy.

Greengrass is more of a cloud first, if you like, in that you would go to the AWS console and configuring the console, and that would then it’d create the package for you. Help you deploy it more indirectly? I would say it’s more abstract. One of the things that we’re looking at, specifically with Cortex A three 20 as well is security.

You mentioned security earlier and of course when you’re deploying a diverse set of devices in a [00:23:00]complex distributed system, and then you are connecting those devices both locally. Into a cloud environment such as AWS it’s incredibly important that those devices are kept secure. Do you have any particular security requirements that you place on devices that are interacting with a WS cloud?

Yeah, so I, I would emphasize or expand that a little bit and say. Regulatory compliance in the IO OT space is growingly complex as well. So we, we have like minimum requirements in that, the TLS version you are using, everything is mutually authenticated and then there are best practices which we document, but because we don’t actually make the end devices, we can’t enforce, for example, making sure that your.

Security credentials are stored securely and not exposed to the software and that kind of thing. So any connectivity to AWS iot, [00:24:00] AWS’s first priority is always security, but beyond that, our customers are having to comply with regulation, which is. Coming in. One of my desires for the future is that a lot of these regulatory requirements that are coming in actually unify in some way because they’re very fragmented at the moment.

I think there’s, obviously a recognition that security is becoming more important and the regulation is always a way of enforcing that. So helping customers comply with things like data residency, for example, is a huge thing That’s I suppose that isn’t security, but I put it in the same bucket as regulatory.

It certainly it still applies if you’re looking at some things such as eu, CRA coming up, of course, we could do a whole other podcast on a, yeah. CRA, maybe that’s something to add to Brian’s list for the future. I think some of the things that you know the A three 20 provide the pointer authentication and [00:25:00] this kind of thing I think could be very interesting for Greengrass B two, I should say.

At the time of this conversation, we are bringing up. Greengrass V two and Free Altos on a, an arm V nine A. Still a work in progress, but both as, as far as my investigation goes, both should just run as they are now, but then, like I say to really alleviate customers from having to think about these things.

Some of the pointer authentication. Which we’ve built into the Cortex M with in collaboration with arm, I should say Good collaboration going on there. We can now bring into those cortex a applications as well. Very important, again, in iot devices, if people have physical access to devices, then the threat vectors are much larger.

And being able to detect things like people manipulating pointers, is it jump orientated or return orientated attacks and that kind of thing to make use of the features that are in [00:26:00] the RV nine A. To trap these things without the customer having to think about it is again, always helping the customer.

I think. Yeah, and it’s looking at the industries that we’re targeting specifically within the IOT business at arm. In my view, it’s non-negotiable. And of course there are always two elements to this. You have the architectural capabilities that we’re building into the ip, and I may be sharing my bias here, but you also have to expose those capabilities in a way that they can be consumed.

And so of course. Making sure that in addition to those things you mentioned, we are bringing in new architectural capabilities into the A three 20, such as pointer authentication, branch target identification as well, where you’re essentially signing pointers to make sure that they’re valid to prevent the kind of.

Threat vectors that you were mentioning earlier, and both of these are examples of popular attack vectors that become far more difficult if those features are being implemented correctly. And previously [00:27:00] we hadn’t seen this widely deployed in IOTs devices. So I think the fact that’s natively supported in the architecture is a really important step forward.

And in addition to us providing those architectural capabilities, of course, we also have to provide a base firmware stack that helps developers to make use of it as well. So not just with the Cortex a three 20, but also with the other IP that we bring to market. We make sure that includes things such as the trusted firmware stack, which enables a secure boot flow.

For example, we also provide a an opti implementation, which provides a runtime for things such as key storage and secure crypto. And I think the thing to really look at here is that these are all tools that can be used, and there’s not one perfect picture. It depends on the implementation. It depends on the markets that our customers, that our partners gonna market [00:28:00] are looking to target, but at least providing that consistency with all of our ip.

So in each of these cases, we’re thinking not just about the architectural capabilities that are being delivered, but we’re also thinking about how we expose those capabilities to consumers of that platform so that everyone can take advantage of them as well. Earlier I said that we don’t actually make the devices so we can advise people what to do, but we don’t actually see what they do and am I right in saying that, lots of this pointer authentication is just a compiler flag and.

And then you have it. So that’s a big help to customers. If you think about how customers get things wrong, a lots of these security things, like I say are quite nuanced, quite mathematical and require additional configuration. And to be able to, they just say, switch this compiler flag on, and then you are getting all these benefits is a real step forward in usability and.

Everyone really protecting customers from themselves, even if they get [00:29:00] these things wrong. Yeah, absolutely. I think there are, you mentioned non-differentiating before, and that’s a balance that we have to be wary of at arm with a lot of things that we do because each partner’s version of non-differentiating maybe slightly different, so we can’t necessarily.

Say to partners when they’re bringing device to market, they have to support things such secure boot or they have to support such as U update. Our job regulation may enforce some of that in the near future, but that’s not opposition to do that. But what we can do is we can provide all of the mechanisms that make sure that when that regulation comes into effect, or when licensees of that IP see use cases or have customers that do require that functionality, it’s as easy as possible for them to implement it if they have to do something at all.

In fact, in the example you just mentioned. You probably have to disable it rather than gonna enable that flag in the compiler. So it is a good example of how we’re raising the bar, [00:30:00] and I think that’s, it’s really important that we think about it in that way. A security’s never going to be a solved problem, but making sure that people have the tools available to them, making sure that we’re raising the bar as much as possible and making it easy to deploy those capabilities is really important.

Yeah, it’s all layered, isn’t it? Layered, it’s, yeah. Security and zero trust and all that. So anything that enables that. We spoke a bit of the, during the intro about machine learning and ai. I’m interested in understanding how Greengrass supports. Different machine learning capabilities as well, the kind of components that you see.

I know you said earlier that you’re providing a generic set of capabilities, but understanding how those are consumed and how those tie into some of the use cases that everyone in the industry at the moment is talking about would be really interesting to understand. Yeah. It’s unop opinionated again in a sense.

So how are you, manage the orchestration of different models, especially if they are tuned to different [00:31:00] immediate environments. By which I mean if there’s a cardinal cardinality problem in that if you are measuring vibration on a machine, then even the same machine in a different environment can have different characteristics.

How solid is the flaw, for example? That the machine is running on, but what’s the acoustic environment that it’s running in. So how you manage the kind of cardinality of fine tuning models and deploying them to the correct devices. It can help a lot. There. Use cases where we see.

Of growing usage is as there is more, more compute done at the edge and more data involved, always more data, then not allowing a cloud connection to become a bottleneck or just reducing the cost of the cloud connection. Growing number of use cases where ML is, you can store things locally on Greengrass and then ML is used to look at a large data set and then just select the pieces, which are.[00:32:00]

And only send that to the cloud. So that’s another growing use case. Then as I was saying before anything which is kind of Docker based, it, it can help you with as well. Plus reporting back. Telemetry and knowing how your model is performing through that telemetry. It’s the plumbing around the application that’s using ml.

Did that answer your question? I’m not quite sure That did answer my question. I have one more question though. I think I’d be really interested to get your view, because you have such a wide view of the industry and it would be really interesting to understand what as the future trends. Coming.

Some use cases that you are beginning to see that either you didn’t expect or that, that are making you stop and think a little bit. What, where do you think we’re gonna be in a few years time as we’re speaking about all of this capability that we’re bringing in? No, that’s a in interesting conversation.

I, I suppose I’ve just alluded to that a little bit already in the when I was talking about the way, ML is being used more and [00:33:00] more. It has been for a while, but more and more in this, allowing you to store more at the edge and still get the benefit from sending data the cloud.

So here. Age, as I alluded to before, where there are some cyclical trends that you see. So we can extrapolate and say those trends will continue. And in particular, how compute moves between accessing remote compute and local computes. So if you think of where we have, going back a bit, mainframe computers and how that moved.

As technology progressed and the cost came down, and then we had, desktop computers and then with the cloud, the the scalability and the pays you go model and all those things. And just a huge compute that things moved back to the cloud. And then especially in the real time, we’re coming back to, as we’ve been talking about in this podcast in this conversation, how we’re doing more at the edge.

Again, we just think, we can almost predict and say what’s the big growing cloud workloads at the [00:34:00]moment? Then we can see, okay, there’s AI and our large language models in the last few months, and particularly multimodal models, just how that’s evolving and say, okay.

Technology evolves, and that’s gonna move to the edge, for all the normal reasons, closer to the data, make faster decisions, latency, data sovereignty, all the things that are driving that trend already. And then, we have devices or cause the the Cortex a three 20, that are gonna, we’re starting to see how we can move these models to the edge more and more already.

I would say one of the. If you try and think about what that’s going to look like or what the result of that is going to be, then I think how can we use multimodal models at the edge? Low power. So let’s go really low power and think about wearables, medical devices, fitness trackers, if they are able to.

Perceive what’s going on around them. Maybe there’s a camera, which is [00:35:00] looking at what you’re eating, maybe what you’re drinking even, and then it’s measuring your calorie. Burn rate because you are walking or cycling or what your heart rate’s doing, and then give you a personalized fitness plan, either a dietary plan or how much exercise you need to maintain a weight or a goal or whatever it is you’re trying to do.

Now, today, you can do that of course, but you have to enter all this information manually and people cheat, of course. So I think, that’s really where I see things going and I don’t think it’s going to be that long either. I’m inclined to agree. I think the seeing more intelligence distills into devices with very small power envelopes, and also you alluded to it in your description, but the context that’s required in order to make more intelligent reasoning decisions whilst taking in that input.

So rather than just. Taking in data and then saying in order for this to be processed, we’ll have to send it to the cloud, and then we’ll get the result back. I think having a [00:36:00] greater level of intelligence at the edge and being able to take in context of your surroundings will be really fascinating.

Although I hope I won’t be judged too harshly for my breakfast in the morning. Bye. Whatever’s looking at my breakfast plate. The other way that can go, and we are seeing this a little bit, is clusters of devices. So having low power devices which are able to perform one piece of perception somehow, but directly communicating you.

Getting a broader picture, a greater situational awareness by directly communicating with other devices that have another view of the world and building up a picture like that again, without a cloud hop or using the cloud when there is something significance in there. In there. Situational awareness.

Yeah. That’s fascinating. That’s really interesting. I think this is a good opportunity too. Thank you for your time. We could talk about this for hours and hours, but this has already [00:37:00] been really interesting, so thank you so much for coming to speak with us and thank you for, thanks for your contributions.

Thank you. No worries.

Brian: Yeah, gentlemen, it’s been like listening to two great musicians talk about their craft, so from the sidelines, I add my thanks. Thanks, Brian.