1
00:00:07,490 -->
00:00:07,880
[music]
2
00:00:07,880 -->
00:00:18,200
Hello there, everyone. Welcome to episode number
701 of this here electronic engineering podcast
3
00:00:18,200 -->
00:00:26,960
called Amelia's Weekly Fish Fry, brought to
you by eejournal.com and written, produced,
4
00:00:26,960 -->
00:00:33,840
and hosted by yours truly, Amelia Dalton. In
this week's podcast, we're diving deep into
5
00:00:33,840 -->
00:00:40,840
the frontier of artificial intelligence and
exploring how mimicking the human brain could
6
00:00:40,840 -->
00:00:49,760
hold the key to ultra-low power, highly efficient
computing at the edge. My guest is Sean Hehir,
7
00:00:49,760 -->
00:00:57,440
CEO of BrainChip. Sean and I explore their
groundbreaking Akida neuromorphic processor
8
00:00:57,440 -->
00:01:05,360
and how BrainChip's event-based approach
processes data only when meaningful changes occur,
9
00:01:05,360 -->
00:01:12,120
shifting away from traditional, power-hungry
accelerators and paving the way for the future
10
00:01:12,120 -->
00:01:20,320
of always-on smart technology. So, without
further ado, please welcome Sean to Fish Fry.
11
00:01:21,120 -->
00:01:26,760
Hi, Sean. Thank you so much for joining me. Hi,
Amelia. Great to talk. Excellent. All right. So,
12
00:01:26,760 -->
00:01:32,720
before we dive into the technology, can you give
us an overview of BrainChip and its vision for
13
00:01:32,720 -->
00:01:41,000
bringing ultra-low power, event-based AI to the
edge? And what makes your neuromorphic approach
14
00:01:41,000 -->
00:01:48,120
fundamentally different from conventional
AI processors? Sure. There's three questions
15
00:01:48,120 -->
00:01:52,880
in there, so let me start with the first one, a
little bit about what was what drove us to create
16
00:01:52,880 -->
00:01:58,120
this company or the vision for this. The company
was founded by two founders. I'm not a founder.
17
00:01:58,120 -->
00:02:04,040
I was brought in about 4 and 1/2 years ago as a
CEO. But one of the founders, Peter Van der Meer,
18
00:02:04,040 -->
00:02:10,880
had this vision of creating computational devices
that mimic the brain because the brain is the most
19
00:02:10,880 -->
00:02:15,720
power efficient computation engine in the world,
right? If you think about what's going on with
20
00:02:15,720 -->
00:02:20,480
you and I just having a discussion right now,
we're utilizing power, our brains are working.
21
00:02:20,480 -->
00:02:24,280
But it's very efficient. And the
point about your brain, it ignores
22
00:02:25,000 -->
00:02:30,840
things that don't matter. And like the example
I like to give a lot is when we meet people,
23
00:02:31,360 -->
00:02:35,640
if you meet a new person, you always look at the
differences. Most human beings have, you know,
24
00:02:35,640 -->
00:02:41,320
similar features. We all you typically have two
eyes, nose, mouth, ears. We look for differences.
25
00:02:41,320 -->
00:02:48,240
Hair color, glasses, facial hair. And so, similar
to that, if you think about what our company does,
26
00:02:48,240 -->
00:02:53,520
we have a type of processing that this ignores
what we call non-events. If nothing's happening,
27
00:02:53,520 -->
00:02:57,680
you ignore it, which makes it inherently very,
very efficient. And I know we're going to talk a
28
00:02:57,680 -->
00:03:02,280
little bit more about the technology, you know,
later. So, maybe I'll just leave it at that.
29
00:03:02,280 -->
00:03:07,680
That was the vision for the company first and
foremost. The second question was more around
30
00:03:07,680 -->
00:03:12,800
repeated for me, if you don't mind. Sure. What
makes your neuromorphic approach fundamentally
31
00:03:12,800 -->
00:03:18,560
different from conventional AI processors? Sure.
Well, it's similar in nature. So, the way we look
32
00:03:18,560 -->
00:03:25,360
about our approach is most conventional
computational approaches are brute force
33
00:03:25,360 -->
00:03:31,720
multiplication, right? You get a certain number of
X by Y axis and you're multiplying and multiplying
34
00:03:31,720 -->
00:03:36,880
and multiplying, whether anything's happening
or not, or if there's any events in there. So,
35
00:03:36,880 -->
00:03:41,160
when you're an event-based or
neuromorphic event kind of technology,
36
00:03:41,160 -->
00:03:46,080
if nothing is happening, it is a zero in that
middle of that matrix versus any other value,
37
00:03:46,080 -->
00:03:50,760
we don't multiply. And if there's a data stream
and nothing's happening, we don't do anything. We
38
00:03:50,760 -->
00:03:55,760
just sit there in idle and wait. So, that's really
the fundamental difference right there. I know
39
00:03:55,760 -->
00:03:59,560
there's a follow-on question or I believe you're
going to ask me later about other attributes,
40
00:03:59,560 -->
00:04:04,160
but that's the core of it. And the interesting
thing about that and when you think about it,
41
00:04:04,160 -->
00:04:08,640
it makes perfect common sense as a paradigm for
those of us who've been in the industry for a long
42
00:04:08,640 -->
00:04:14,240
time and it's kind of brute multiplication.
It was just left over from 40-50 years ago,
43
00:04:14,240 -->
00:04:20,080
early parts of computation or computing, but it
doesn't make sense. This is the logical approach.
44
00:04:20,080 -->
00:04:25,280
Only compute when something's needed, otherwise do
nothing. That's right. Okay, so let's talk about
45
00:04:25,280 -->
00:04:32,120
Akida. BrainChip's neuromorphic processor built on
those spiking neural networks. Now, traditional AI
46
00:04:32,120 -->
00:04:41,200
accelerators constantly process data streams while
Akida uses event-driven neuromorphic approach. So,
47
00:04:41,200 -->
00:04:48,760
can you explain how spiking neural networks
allow Akida to compute only when something
48
00:04:48,760 -->
00:04:53,920
meaningful changes? Sure. First of all, let
me even clarify the statement cuz you know,
49
00:04:53,920 -->
00:04:59,640
what you try to do when you get a breakthrough
technology like us, you don't want to make it
50
00:04:59,640 -->
00:05:05,520
hard to deploy, right? Because when people hear
the word spiking or SNNs, they kind of maybe say,
51
00:05:05,520 -->
00:05:10,280
"Oh, this is scary, you know, it's something
different." What we do is specifically, we get
52
00:05:10,280 -->
00:05:14,800
this breakthrough performance that, you know,
I highlighted on and we'll talk further about,
53
00:05:14,800 -->
00:05:19,160
but we make it easy to adopt traditional
models. So, if you've got a traditional model,
54
00:05:19,160 -->
00:05:24,280
we've got a tool chain and you run it through,
it makes it Akida compatible and it allows you
55
00:05:24,280 -->
00:05:30,120
to act like an SNN if you will in that regard. So,
you've got to get that kind of thing. Now, I think
56
00:05:30,120 -->
00:05:34,960
I already kind of gave the answer to it a a moment
ago at that level. What makes it so efficient is
57
00:05:34,960 -->
00:05:40,240
the data streaming. When there's a zero, we don't
do anything. It's just that simple. So, we take it
58
00:05:40,240 -->
00:05:46,240
and if nothing's happening, it's perfect. So, the
ideal use case to say is where sparsity is high.
59
00:05:46,240 -->
00:05:51,440
The more sparsity, the more power efficient
that is. So, anything when you go through this,
60
00:05:51,440 -->
00:05:55,920
wearable devices that you monitor when nothing
is going on, that I could give you a lot of
61
00:05:55,920 -->
00:06:00,320
examples about that. Preventative maintenance
where nothing's going on, but when it goes on,
62
00:06:00,320 -->
00:06:04,920
something happens. In the state of the
world we are in today in defense industries,
63
00:06:04,920 -->
00:06:10,920
the ability to do these things and monitor large
amounts of data streams when nothing's going on,
64
00:06:10,920 -->
00:06:15,040
but if something's happened, you need to activate
and do something quickly. These are ideal these
65
00:06:15,040 -->
00:06:19,600
cases for the kind of technology we have. So,
talk to me a little bit more about what's going
66
00:06:19,600 -->
00:06:25,800
on under the hood in Akida. How does it reduce its
power consumption to the microwatt to milliwatt
67
00:06:25,800 -->
00:06:32,280
range instead of constantly running like a
traditional processor? Yeah, it's it's really the
68
00:06:32,280 -->
00:06:37,600
fact that I highlighted earlier. There's really
three main factors, but the primary one is just
69
00:06:37,600 -->
00:06:42,560
doing nothing when things aren't happening.
You got to remember most neural networks today,
70
00:06:42,560 -->
00:06:49,320
if 50% of them are really have non-activity going
on in a particular point. In the traditional way,
71
00:06:49,320 -->
00:06:53,160
you're going to use that computation and
compute on something that's not going on.
72
00:06:53,160 -->
00:06:56,760
So, you take out 50, you know, roughly
50%. Some models are going to be higher,
73
00:06:56,760 -->
00:07:01,680
some are going to be lower. You take that
out, that's an enormous savings right there.
74
00:07:01,680 -->
00:07:08,520
Secondly, we have a world-class quantization kind
of technology that allows you to reduce the data
75
00:07:08,520 -->
00:07:13,920
and not lose any accuracy or minimal accuracy. So,
if you reduce the amount, you know, weights and
76
00:07:13,920 -->
00:07:19,720
activity going on, you're going to have massive
power savings. And the last big piece about Akida
77
00:07:19,720 -->
00:07:25,680
when you put it all together is that that near
memory compute. And so, the way we structure this,
78
00:07:25,680 -->
00:07:30,000
we have what we call node structures.
Each node has a single compute engine
79
00:07:30,000 -->
00:07:35,600
and four little pieces of memory around it. And
depending on how big of a chip you want to build,
80
00:07:35,600 -->
00:07:41,240
or you know, we have our own chips, but if we're
selling IP, you stack out the number of nodes, and
81
00:07:41,240 -->
00:07:46,360
once you load the weights into that memory, they
stay in the activity. So, you're not moving things
82
00:07:46,360 -->
00:07:52,480
on and off the chip. You put that together, you
get a very low power usage for your technology.
83
00:07:53,320 -->
00:08:00,440
Absolutely. Okay, so efficiency involves more than
just reducing the number of computations, right?
84
00:08:00,440 -->
00:08:08,240
So, how does Akida reduce the overall energy cost
of AI at the edge? If you're not using the power,
85
00:08:08,240 -->
00:08:12,240
you're not using the energy. And usually
And usually when people use the word edge,
86
00:08:12,880 -->
00:08:18,800
usually there's a battery, right? If there's
a battery, then lower power usage allows you
87
00:08:18,800 -->
00:08:24,880
to extend that battery life and make use cases
that were not possible or even dreamable. And
88
00:08:24,880 -->
00:08:29,240
I like examples, they bring them to life. We
work with a company in a different part of the
89
00:08:29,240 -->
00:08:37,200
world that is making a set of glasses with with a
sensor on it that can pick up EEG signals to say,
90
00:08:37,200 -->
00:08:42,160
"You are about to have an epileptic seizure." If
you happen to have epilepsy. And what that does,
91
00:08:42,160 -->
00:08:46,000
if you have that, it says, "Okay,
the signal comes into the sensor.
92
00:08:46,680 -->
00:08:52,320
We we analyze and says, "Yes, this has the
pattern of a potential." They claim that it's
93
00:08:52,320 -->
00:08:57,920
98% accurate, their model right now, that you will
have a seizure within 1 hour of that prediction.
94
00:08:58,600 -->
00:09:04,320
That's a powerful claim. Now, for people who
suffer from that challenge, it allows a freedom
95
00:09:04,320 -->
00:09:08,960
in their lifestyle they would they wouldn't. They
can now drive. You know, stories I've met people
96
00:09:08,960 -->
00:09:12,480
when this product was being developed that
were afraid to leave their home. They can now
97
00:09:12,480 -->
00:09:18,560
have some freedom. So, these use cases like that
were not possible. We have a licensee for our IP
98
00:09:19,320 -->
00:09:24,080
in Europe that's building something for space.
Cuz space is a lot of times you're going to lose
99
00:09:24,080 -->
00:09:28,440
your network. You need to do inference locally.
You need battery life. It's space, you're not
100
00:09:28,440 -->
00:09:33,960
going to plug anything in. Ideal for those use
cases. It's not only isolated to battery or
101
00:09:33,960 -->
00:09:39,520
ones. We also have where power cords matter. But
those kind of cases, wearable devices, you know,
102
00:09:39,520 -->
00:09:44,120
space, prevent and maintenance. You know, we're
working with certain companies that have remote
103
00:09:44,120 -->
00:09:48,720
locations for preventive maintenance on oil rigs
and things like that. And then, of course, the
104
00:09:48,720 -->
00:09:56,040
I mentioned earlier, defense industry, everything
is mobile right now. Drones, robots, those matter.
105
00:09:56,040 -->
00:10:01,560
Quick action, low latency security, you don't want
data moving about, and low power. All those things
106
00:10:01,560 -->
00:10:07,840
matter. Ideal use cases for those. So, Sean,
finally, what does this architecture change
107
00:10:07,840 -->
00:10:14,480
about the future of edge AI, do you think? Where
do you see neuromorphic processors, like Akida,
108
00:10:14,480 -->
00:10:22,560
having the greatest impact? And what applications
weren't practical before because conventional AI
109
00:10:22,560 -->
00:10:28,200
just required too much power? I think it changes
everything, right? You know, first of all,
110
00:10:28,200 -->
00:10:34,400
changes in technology are usually generational
with people, right? You got to always remember,
111
00:10:34,400 -->
00:10:41,240
arguably, better technology always doesn't prevail
if people won't adopt it. Often in technology,
112
00:10:41,240 -->
00:10:47,200
engineers that have been trained 20 and 30 years
ago don't want to consider new technologies.
113
00:10:47,200 -->
00:10:53,160
Younger generations of engineers now recognize the
potential. And [clears throat] so, we're seeing
114
00:10:53,160 -->
00:10:57,520
that. If you look at the market forecast for these
kind of technologies, neuromorphic technologies in
115
00:10:57,520 -->
00:11:03,040
edge, it's up into the right and very rapidly
cuz it's just too compelling to keep Why not?
116
00:11:03,560 -->
00:11:07,680
You know, you asked the question about what
does it open up? I gave you some examples.
117
00:11:07,680 -->
00:11:10,880
I'll give you a couple more, but I'd
like to answer the question this way.
118
00:11:10,880 -->
00:11:15,040
I became the CEO of this company 4 and
1/2 years ago. When I first came here,
119
00:11:15,040 -->
00:11:20,280
my customer conversations were similar to
the one we're having. Interesting technology,
120
00:11:20,280 -->
00:11:26,120
help me envision what I can do with it. That's not
what happens today. Today, the conversations are
121
00:11:26,120 -->
00:11:31,920
Shawn, or whoever in our organization, I am
building this product. I have these requirements.
122
00:11:32,560 -->
00:11:36,440
You can do this. I don't think anybody
else can. Help me build this with this.
123
00:11:36,440 -->
00:11:40,920
That's the power what's going on the market
now because people see the potential.
124
00:11:40,920 -->
00:11:45,480
I don't know if you go to CES or these other shows
like that. You see these new products are coming
125
00:11:45,480 -->
00:11:52,200
out that required these You cannot do this if you
were really concerned about low power, security,
126
00:11:52,200 -->
00:11:57,680
latency, any use case. I just gave you a bunch,
you know. Wearable devices, defense, preventive
127
00:11:57,680 -->
00:12:03,120
maintenance. They're everywhere. Medical. And
so, it opens to a whole new class of these
128
00:12:03,120 -->
00:12:06,800
You know, and they're not going to give you
even more contextual conversation. Probably not
129
00:12:06,800 -->
00:12:11,280
the essence of your question, but I I still think
it's worth mentioning. Most people think AI today,
130
00:12:11,280 -->
00:12:14,840
if you talk to your neighbors, think of the
classic data center, ChatGPT, and things like
131
00:12:14,840 -->
00:12:19,440
that. And that's fine, and it has its place. But
we also, at the same token, if you were at your
132
00:12:19,440 -->
00:12:23,520
neighbor's barbecue, they would say, "Well, these
data centers and where are you even using power?"
133
00:12:23,520 -->
00:12:29,480
The world has got to think of ways to do it
smarter. We can't keep doing this. Now, again, I'm
134
00:12:29,480 -->
00:12:35,640
an industry veteran for many decades in computing.
Then everything starts in a centralized model like
135
00:12:35,640 -->
00:12:41,920
AI did. And then you decentralize, and you get
to this distributed hybrid computing. AGI is in
136
00:12:41,920 -->
00:12:47,200
that growth phase right now. And what if you came
5 years from now we had this conversation, and if
137
00:12:47,200 -->
00:12:51,160
you we had a way to quantify the measurement, how
much is in a data center, how much is in an edge,
138
00:12:51,160 -->
00:12:55,920
you're going to see a lot of that work move out
because you don't have to do it that way. And
139
00:12:55,920 -->
00:13:00,200
also, people want more control. Now, I would also
say the following. I think it's really important.
140
00:13:00,200 -->
00:13:04,400
When you build products, because you know, in the
end, we're enabling revenue streams for companies.
141
00:13:05,200 -->
00:13:10,800
They're going to compete on performance against
their own competitors. So, they want the low power
142
00:13:10,800 -->
00:13:15,680
for things like I could give you so many examples
saying My one of my first things I talk about to
143
00:13:15,680 -->
00:13:20,480
companies is who's your competitor? What are
you worried about? And they always require
144
00:13:20,480 -->
00:13:24,640
the things that we deliver to deliver the
functionality for their end products. So,
145
00:13:24,640 -->
00:13:28,600
I think it it everything the answer to case. And
I think And every day, one other and I'll be brief
146
00:13:28,600 -->
00:13:32,880
and I know I'm going a little long-winded on this,
but it's important. I give you some examples, but
147
00:13:32,880 -->
00:13:38,320
I'm constantly energized by the examples I don't
think of or we think of and our customers say,
148
00:13:38,320 -->
00:13:43,320
"I want to do this." And I'm like, "Wow, that's
a great use." So, every year we learn a lot more.
149
00:13:43,920 -->
00:13:49,600
I love that. That is fantastic. All right, Sean,
it's time for your off-the-cuff question. So,
150
00:13:49,600 -->
00:13:54,640
if you get out one meal right now, doesn't matter
if it's on the other side of the world, you need a
151
00:13:54,640 -->
00:13:59,520
passport to get there, what would you have? I'm
going to go with my comfort food. I travel the
152
00:13:59,520 -->
00:14:04,520
world, by the way, and I love all kinds of exotic
food, but if I had one meal, my comfort food,
153
00:14:04,520 -->
00:14:10,720
my doctor won't like this, I love fried
chicken sandwiches and iced tea. I'm sorry,
154
00:14:10,720 -->
00:14:15,760
not sophisticated to answer, that's the truth. I
love that. That sounds absolutely fabulous right
155
00:14:15,760 -->
00:14:21,400
now. Well, Sean, this was super cool. Thank you
so much for joining me. Thank you, Amelia. It
156
00:14:21,400 -->
00:14:25,200
was a real pleasure. Look forward to talking
again in the future. Have you had a chance to
157
00:14:25,200 -->
00:14:33,840
check out last week's 700th episode of this here
podcast? In honor of such a momentous occasion,
158
00:14:33,840 -->
00:14:41,160
I reported on the tiniest of technologies. From
light-powered itty-bitty robots that can hunt
159
00:14:41,160 -->
00:14:51,000
down and collect bacteria to new memristors built
using synthetic DNA-based biohybrid architectures
160
00:14:51,000 -->
00:14:57,120
to new nanolasers that could make
next-generation computers far more efficient.
161
00:14:57,640 -->
00:15:02,960
And you can check out this episode by clicking
the link below the player on this week's Fish
162
00:15:02,960 -->
00:15:13,760
Fryin' page on eejournal.com or by heading on
over to YouTube, youtube.com/eejournal. Also,
163
00:15:13,760 -->
00:15:20,200
if you'd like further information about BrainChip,
I've also included a couple links below the player
164
00:15:20,200 -->
00:15:27,440
on this week's landing page on eejournal.com
and in the description for this week's YouTube
165
00:15:27,440 -->
00:15:35,800
episode as well. Hey, have you checked out ee
journal on social media yet? Well, you should.
166
00:15:35,800 -->
00:15:43,320
You can find us at facebook.com/eejournal.
If LinkedIn is more your thing, I dig it.
167
00:15:43,320 -->
00:15:51,920
You can follow us or me on LinkedIn and we are
also on BlueSky Social and Mastodon as well.
168
00:15:51,920 -->
00:16:01,200
And we have that YouTube channel I just mentioned,
youtube.com/eejournal. Folks, it is chock-full of
169
00:16:01,200 -->
00:16:09,520
all kinds of techie videos including our very
popular Chalk Talk webcast series hosted by me.
170
00:16:10,480 -->
00:16:16,200
Thank you everyone for tuning in. If you know of
any cool new technology or heck, you just want to
171
00:16:16,200 -->
00:16:26,840
chat, shoot me a line at Amelia, that's a m e l i
a @ eejournal.com or post a comment on our forums
172
00:16:26,840 -->
00:16:35,760
on EE journal. For the week of October 2nd,
2026, I'm Amelia Dalton and you've been fried.
