Latest / EDGE AI POD / How Microsecond AI Control Transforms Power Systems And Cuts Errors

Transcript

How Microsecond AI Control Transforms Power Systems And Cuts Errors · EDGE AI POD

• 0:06
  SPEAKER_00: Good morning.
• 0:08
  SPEAKER_00: Hope everyone's feeling fresh, not too many sore heads out there.
• 0:12
  SPEAKER_00: I'm Martin McDonald, so I'm a chief operating officer here at WeTech.
• 0:17
  SPEAKER_00: So WeTech is a startup we've been going for around about three years now, uh based in Glasgow in Scotland, in Bonnie, Scotland.
• 0:27
  SPEAKER_00: Uh so we're all struggling with the heat over here in uh in Milan.
• 0:30
  SPEAKER_00: Um so we're gonna be talking about ultra-low latency edge AI and the key to unlocking nonlinear control beyond the edge.
• 0:40
  SPEAKER_00: So I'll take you through this.
• 0:42
  SPEAKER_00: We have a demo uh table downstairs if you go downstairs and turn left.
• 0:46
  SPEAKER_00: We're right in the back right hand side corner.
• 0:49
  SPEAKER_00: Um, if you want to find out some more.
• 0:52
  SPEAKER_00: So you might be sitting there thinking, well, hang on a minute, edge AI uh is already low latency, so what's the problem?
• 1:00
  SPEAKER_00: And you'd be perfectly within your rights to ask that problem.
• 1:02
  SPEAKER_00: There is no problem.
• 1:03
  SPEAKER_00: I would ask you to rephrase the question and ask what's the opportunity.
• 1:08
  SPEAKER_00: So, where we see the opportunity is where we've focused our attention and focused our research is on power and control system optimization.
• 1:18
  SPEAKER_00: So, technology-wise, think things like motor drives, power inverters, energy storage systems, uh, and from an industrial vertical perspective, where do we target?
• 1:28
  SPEAKER_00: Uh, automotive is key, industrial automation is massive, uh, certainly in the motor drive piece.
• 1:34
  SPEAKER_00: Utilities and data center, we're starting to see a lot more activity in there.
• 1:38
  SPEAKER_00: We're starting to see NVIDIA talk about 800 volt uh to the rack data centers of the future, and this is something where we see a real opportunity for this ultra-low latency edge AI.
• 1:50
  SPEAKER_00: So, just talking about system linearity and control in very basic terms, uh, right at the bottom we have uh linear systems, uh, so effectively where the input is directly proportionate to the output.
• 2:04
  SPEAKER_00: Um, and as we move up the scale in terms of linearity, we get to a point where we have nonlinear control.
• 2:11
  SPEAKER_00: So you can think of linear control as something like an electric motor mixing a VAT of some kind of liquid.
• 2:18
  SPEAKER_00: Um, and at the top end you might have a VAT of some kind of rubber for a tire manufacturer, for example, where you can't necessarily predict uh the output from the input.
• 2:29
  SPEAKER_00: And as you step through these different uh layers, if you like, of control, we have different control techniques which have been developed for decades, uh, which we can effectively use very efficiently in linear control systems.
• 2:43
  SPEAKER_00: But as we get through to the higher order and the nonlinear control systems, we start to have to do more complex control scenarios.
• 2:51
  SPEAKER_00: Uh, and this is an issue, this is a challenge.
• 2:54
  SPEAKER_00: Um, you can always throw money at the situation and get to a point where you have real-time control, but that's where the real world kicks in.
• 3:03
  SPEAKER_00: So, in the real world, we have product constraints, which are usually around cost, memory, performance, etc.
• 3:10
  SPEAKER_00: So the possibility of doing real-time control when you get up to this higher end is still there, but you have to throw a lot of money at it.
• 3:18
  SPEAKER_00: So, actually, the feasibility of being able to do real-time control when we get up to higher order and non-linear control is a real challenge.
• 3:26
  SPEAKER_00: And that's the challenge that we're addressing.
• 3:30
  SPEAKER_00: It's a challenge that we faced in the past, and uh, those of a certain age will recognize this transition.
• 3:36
  SPEAKER_00: I certainly do, uh, from voice uh uh applications mainly and text on these older phones that we used to see back in 2007.
• 3:49
  SPEAKER_00: And we transitioned to a point where we had the iPhone and the smartphone that we all know, love, and hate at the same time on the right hand side.
• 3:57
  SPEAKER_00: But can you tell me how long it took for that transition?
• 4:01
  SPEAKER_00: If anyone wants to offer a suggestion in number of years, I'll put myself out of my misery in a minute if you can't think.
• 4:11
  SPEAKER_00: One year.
• 4:13
  SPEAKER_00: So between 2007 and 2008, this is how mobile phone technology changed.
• 4:20
  SPEAKER_00: And we could do things like build big bigger batteries, of course, um, but it's still got to fit in your pocket.
• 4:26
  SPEAKER_00: So the thing that was introduced at that point was this dynamic voltage frequency scaling.
• 4:31
  SPEAKER_00: So changing the frequency based on the dynamic input, what apps you're using, etc.
• 4:35
  SPEAKER_00: And I tell you this for two reasons.
• 4:37
  SPEAKER_00: I tell you this first of all because this is a challenge that industry has solved.
• 4:41
  SPEAKER_00: And I for the second reason is that my co-founder and colleague Tana Dossololo over here was one of the first engineers to work on this problem almost 20 years ago.
• 4:50
  SPEAKER_00: So this was a paradigm shift really in power management capabilities in the mobile phone industry, and we're going through a similar uh paradigm shift right now.
• 5:00
  SPEAKER_00: So, what we do as a business is we we do system optimization, uh, but this is an idea of shifting system optimization of being the domain of the cloud, local, and even edge, uh, to one where we're doing dynamic system optimization at a microcontroller level.
• 5:17
  SPEAKER_00: So the underlying technology that we have is called ultra-edge, and you can think of this as the way of being able to detect and respond to error signals in less than one-tenth of a millisecond.
• 5:27
  SPEAKER_00: So ultra low latency.
• 5:29
  SPEAKER_00: So a hundred microseconds, we're able to detect and respond.
• 5:32
  SPEAKER_00: But the predict piece is really important.
• 5:35
  SPEAKER_00: We don't need historical training data to do this.
• 5:38
  SPEAKER_00: We get all of the data that we need from the startup of the machine, and we're able to predict the path of an error signal and effectively intervene.
• 5:46
  SPEAKER_00: If we see a more efficient path, we take it.
• 5:49
  SPEAKER_00: And that's effectively the underlying technology that glues all this stuff together.
• 5:54
  SPEAKER_00: So if we look at our uh diagram again, when we deploy uh ultra-edge technology, we can effectively deploy this across both linear systems and nonlinear systems.
• 6:06
  SPEAKER_00: And if you imagine how quick that we can respond, we can start to effectively extend the range of PID control through to nonlinear applications.
• 6:16
  SPEAKER_00: So, meaning that we can do real-time control uh across the whole range.
• 6:23
  SPEAKER_00: So, where do we use this practically?
• 6:26
  SPEAKER_00: This is um the motor control um use case, and you can see this downstairs.
• 6:31
  SPEAKER_00: We have this live.
• 6:32
  SPEAKER_00: Um, so when we talk about this paradigm shift, we talk about an old paradigm, and it's relatively unfair.
• 6:38
  SPEAKER_00: This is really the IoT approach, this is what we're doing now.
• 6:42
  SPEAKER_00: We've been doing this for quite some time.
• 6:44
  SPEAKER_00: Um, and what we do is we want to know more about a machine, we surround it with sensors and network infrastructure, we take that data, which is observed data, we pass it back up the chain for analysis, but we need historical training data to be able to validate that data.
• 7:00
  SPEAKER_00: And then we send some instruction back to the machine.
• 7:03
  SPEAKER_00: This has worked well for probably a decade now, if not longer, but it's not what we do.
• 7:07
  SPEAKER_00: We do something different.
• 7:08
  SPEAKER_00: So instead of deploying more and more sensor and network infrastructure, we're accessing existing operational data directly from the circuit of the device, and this is what we call ultra-edge.
• 7:19
  SPEAKER_00: So we can look for patterns and anomalies in that data at an ultra-early level and use that data so we can send that data back up the chain to predictive maintenance condition monitoring applications.
• 7:30
  SPEAKER_00: But the real clever thing when we talked about the fact that we can do this ultra-low latency optimization is that as well as identifying where we have anomalies and flagging that, we can actually change the output of the control loop in real time.
• 7:44
  SPEAKER_00: So, this is this idea of identifying and correcting inefficiencies literally as they're happening.
• 7:50
  SPEAKER_00: So, what does that look like?
• 7:52
  SPEAKER_00: Uh you can see this downstairs, but where we have on the left hand side an electric motor going through a talk torque step, a torque transition.
• 7:59
  SPEAKER_00: On the left hand side is without ultra edge, and you can see these peak current overshoots, these speed errors.
• 8:05
  SPEAKER_00: And on the right hand side we have ultra-edge kicking in dynamically.
• 8:09
  SPEAKER_00: Um, so when we kick in, you can see that the speed errors have been significantly reduced.
• 8:15
  SPEAKER_00: This is running on our demo platform, which is 32 megahertz clock.
• 8:20
  SPEAKER_00: When we get our silicon back, which will be in about four or five weeks, this will be running at 320 megahertz clock, and those even small errors will disappear altogether.
• 8:30
  SPEAKER_00: This is another way of looking at it.
• 8:32
  SPEAKER_00: Um, so you can see the path in blue, which would have been taken without Ultra Edge, and you can see the direct path that we now take with Ultra Edge kicking in, and those data sets uh switching between two points in the process.
• 8:48
  SPEAKER_00: So that's motor control.
• 8:50
  SPEAKER_00: But some of you will remember, all of you will probably remember.
• 8:53
  SPEAKER_00: I hope none of you were experiencing this a few months ago when we had the blackouts in Spain and Portugal.
• 9:01
  SPEAKER_00: Um this is really something that we're addressing with our power converter application.
• 9:06
  SPEAKER_00: So in Spain and Portugal, yes, there's a lot of analysis still to go on in terms of uh the root cause of the problem, etc.
• 9:14
  SPEAKER_00: But the the finger of uh blame keeps getting pointed back to renewables.
• 9:19
  SPEAKER_00: And the reason for that is that renewables such as solar, for example, don't carry any uh inertia.
• 9:26
  SPEAKER_00: When they go, they go and they go into protective mode and they shut down.
• 9:30
  SPEAKER_00: So countries are uh are literally curtailing the use of renewables because of this being an issue.
• 9:37
  SPEAKER_00: So potentially the route to net zero is impossible.
• 9:40
  SPEAKER_00: Uh we can't get past this point, but we can.
• 9:43
  SPEAKER_00: So with Ultra Edge and our grid forming power converter, we're able to perform this control algorithm and this handshake between uh the grid, renewables, and battery energy storage systems so that we can create this idea of synthetic inertia.
• 9:58
  SPEAKER_00: So if we see a problem with the grid, instead of shutting down and becoming protective, we offer a boost to the grid.
• 10:04
  SPEAKER_00: So effectively we're an asset to the grid rather than a problem.
• 10:08
  SPEAKER_00: Um so a credible route to net zero uh is achievable.
• 10:12
  SPEAKER_00: And the reason for this is if you think about back in the day when we just had coal fire fire power uh stations, etc., you had um turbines with huge spinning wheels.
• 10:23
  SPEAKER_00: So if something if they see a blip on the grid and they shut down, they don't shut down totally because you've still got this huge spinning wheel which is um creating inertia and keeping everything nice and stable.
• 10:33
  SPEAKER_00: We're now in a position where certainly in in countries like Spain and Portugal where renewables, particularly solar, are becoming 50% of the the power uh supply to certain countries, and we have this instability.
• 10:47
  SPEAKER_00: In the future, with ultra-edge power converters, we're able to use the battery energy storage systems to give stabilization to the grid.
• 10:58
  SPEAKER_00: So, what we're not doing, uh we're we're we're doing all this stuff at a local level and we're executing and keeping things nice and um optimized.
• 11:07
  SPEAKER_00: We're not throwing away that data, don't worry.
• 11:09
  SPEAKER_00: We're doing a lot of data analysis, but a lot of trimming of the data at the very um at the microcontroller level.
• 11:16
  SPEAKER_00: So here we have an example of the data that we collect, so a million data points in in one second.
• 11:22
  SPEAKER_00: Um, what that looks like uh in reality is a huge number of subsets of data which we can take advantage of.
• 11:30
  SPEAKER_00: But what we do is we filter at that microcontroller level.
• 11:33
  SPEAKER_00: So all you're gonna need to report on is the transients that we see that we identify that are important.
• 11:40
  SPEAKER_00: And in this scenario, we have six.
• 11:44
  SPEAKER_00: Um, if there's no transients to um to speak of that we want to we want to know and we want to capture, we don't do anything.
• 11:52
  SPEAKER_00: So that's us, and that's ultra edge, and that's the opportunity.
• 11:55
  SPEAKER_00: In terms of adopting this, um, we can start from a single device uh to prove capability.
• 12:02
  SPEAKER_00: We have pilot projects going on at the moment uh with some key um you know global vendors.
• 12:08
  SPEAKER_00: We can deploy this as a soft IP or as a hardware demo platform to validate what we're doing, which is all available now.
• 12:17
  SPEAKER_00: We're building up many use cases.
• 12:19
  SPEAKER_00: I've shown two or three today, but the use cases keep on building.
• 12:22
  SPEAKER_00: And when we build the use cases, we also build the library of models which you guys can take advantage of.
• 12:28
  SPEAKER_00: And as I say, we have our test chip coming back in uh about around about four or five weeks.
• 12:34
  SPEAKER_00: So we'll be building our demo platform so it'll be a nice, easy platform for you to take advantage of and prove ultra edge working in your own environment.
• 12:42
  SPEAKER_00: So if it's not too crude, place your orders.
• 12:45
  SPEAKER_00: So thank you very much.