Fair enough but my main point was mankind is it's own worse enemy. There was regulation around the new ev technology but Elon Musk started DOGE and got rid of the people who were carrying out the 2000 odd safety investigations of his Tesla cars. Problem fixed. People in power that have narcissistic tendencies, ego power trips or greed can cause a lot of harm. IMO
brainchip.com
A relevant interview regarding Chris Eliasmith of ABR, makes for an interesting read:
It has a few relevant parts such as:
SB: Your group has also come up with Legendre Memory Units to represent time. Listeners may remember that BrainChip have been using this approach as well. Can you talk about what they are and why they’re useful?
CE: So we were really working on, ‘How does the brain represent time?’ But we also came to realize that, well, this is a problem in machine learning. Time series is a massive area of research, and it’s really hard. And people have all kinds of different recurrent networks, such as LSTNs and GRUs and a million variants on each of these. Transformers are now what people are using, where it’s not a recurrent network but it kind of spreads time out in space so you can just process stuff in a feedforward way. So there are all of these different approaches.
We started applying this core dynamical system to these problems. And so the obvious thing to do, I think, was basically: you take that linear system—so this is the thing representing the temporal information—and then you put a non-linear layer afterwards, so you can then manipulate that temporal representation however you want. And you’ll learn that using normal backprop. And that’s what we call the Legendre Memory Unit.
More recently, people have taken that exact same structure and called it a state-space model, for obvious reasons—because basically, having a linear dynamical system and then a non-linear layer, that’s a state-space model. And that’s what BrainChip is using, for instance.
Also,
CE: And so for the last couple of years, that’s what we’ve really been focused on: building a chip that can natively run state-space models, run it extremely efficiently—because it’s specifically designed to do that—and fabricate that chip, and go to customers and start getting into all the many different products that you might be interested in having that in.
So that’s something that we’re really excited about, because we just got the chip back for it at the end of August. We got the chip back, we tested it, and we had it up and running and setting world records in low-power running state-space models within a week and a half.
SB: Talk more about this chip. I’m very interested!
CE: It’s not a neuromorphic chip, in the sense that it’s not event-based. It is a neural accelerator, so it has compute right near the memory. The memory is actually a little bit exotic—it’s called MRAM, so magnetoresistive RAM, which means that it’s non-volatile—so you can load your network on there, you can basically leave it, shut off, and it draws almost no power until you need it—and then it can run the model, which is really cool.
We’re able to do full speech recognition. So it could be sitting here basically typing out everything that I’m saying with about a hundred times less power than other edge hardware that’s available on the market—under 30 milliwatts. We can have it typing out whatever language you’re speaking in.
We can do other things with it, too. We can use it to control a device. So you can basically tell the device what you want to do, using natural language—you don’t have to memorize keywords or key phrases, you just say what you want. We’re also working with customers who want to use it to do things like monitor your biosignals—your heartbeat, your O2, your sweating, you know, anything. You can monitor all that and do on-device inference about the state of the person, warn them that they’re going to have a seizure if it’s EEG, or that they’re having some kind of heart palpitation, or what have you.
And we just started our early access program. So we’re working with customers, getting the hardware in their hands, helping them integrate that into their applications. We’re super excited about what this chip can do. It’s just kind of blowing the competition out of the water from a power and efficiency and performance perspective.
REC: An interesting aside, which is kind of like a reality check for me—a group of us went to speak to a DARPA director not long ago, and we were selling an idea of which one of the primary focuses was that it was basically extremely low power. And she could not care less about the power. She cared mostly about latency. That is the thing. And this was an embodied AI application, Giulia. She said, “I will burn all the power that I need to if you can get me the speed and the decisions to happen as quickly as possible, and as effectively as possible.” Which for me was like, “What!?” I expected that the power efficiency argument would’ve been a winner on the day.
It was interesting to me. We were trying to argue for small drones with small brains and so on, and she was like, eh, but what’s the latency? When do you make the decision? Anyway, very interesting.
A separate thing I don't recall people mentioning was the Youtube presentation he did a few weeks ago in which he went into his new chip in more detail. There were a few things worth noting such as listing some specs (eg. the chip uses 22nm node process, and a comparison with competitors:
View attachment 93607
A few takeaways:
- This gives further confirmation that ABR's LMU and Brainchip's State Space models are very similar competitive technologies.
- The rush to get Akida Cloud was likely partly in response to the threat ABR proposed. Akida's Gen 2 chip is still in production. However, while ABR have a chip now,Akida's Gen 2 chip is still in production.from 5th August (see link below), slightly earlier than ABR's chip. ABR got their chip back at the end of August, then had a few weeks of testing plus logistics before it could be provided to customers. Akida 2 would likely have been the FPGA prototype, so probably not the final version with maximum efficiency, but it allowed people to test at least a month earlier than ABR. In general, any new prototype chip can be released to customers this way, which provides an edge.
- In the Youtube clip, ABR were comparing their new chip to the 1st gen Akida and saying it wasn't comparable. It couldn't do more complex tasks like automatic speech recognition. I would guess Gen 2 woul have similar efficiency / performance to ABR's new chip unless there are significant improvements that Brainchip have made to hardware or algorithm efficiency (patent pending). Both chips are being manufactured in 22nm from memory so performance should be comparable when these values eventually become available.
- It's probably no surprise that ABR are working with customers to implement solutions. However, even if ABR have better hardware it doesn't mean customers will rush to adopt it over Akida Gen 2. Brainchip has spent many years building up their ecosystem and working with customers and has other advantages of it's own. Furthermore, Brainchip is more focused on IP than chips, which means they are working with some diffferent customers
- Brainchip's competitive advantage may not be as high as some people think. Akida Gen 3 will probably be important to ensuring they retain this
LAGUNA HILLS, Calif. – Aug 5th, 2025 – BrainChip Holdings Ltd (ASX: BRN, OTCQX: BRCHF, ADR: BCHPY), the world’s first commercial producer of ultra-low power, fully digital, event-based neuromorphic AI, today announced launch of the BrainChip Developer Akida Cloud, a new cloud-based access point to multiple generations and configurations of the company’s Akida™ neuromorphic technology. The initial Developer Cloud release will feature the latest version of Akida’s 2nd generation technology,BrainChip Launches Akida Cloud for Instant Neuromorphic Access
BrainChip launches Akida Cloud, providing instant access to the latest neuromorphic technology for developers and edge AI innovators.
TL;DR — They’re complementary, not direct substitutes Bottom Line: Not “one is strictly better” — they target different ends of the edge-AI spectrumA relevant interview regarding Chris Eliasmith of ABR, makes for an interesting read:
It has a few relevant parts such as:
SB: Your group has also come up with Legendre Memory Units to represent time. Listeners may remember that BrainChip have been using this approach as well. Can you talk about what they are and why they’re useful?
CE: So we were really working on, ‘How does the brain represent time?’ But we also came to realize that, well, this is a problem in machine learning. Time series is a massive area of research, and it’s really hard. And people have all kinds of different recurrent networks, such as LSTNs and GRUs and a million variants on each of these. Transformers are now what people are using, where it’s not a recurrent network but it kind of spreads time out in space so you can just process stuff in a feedforward way. So there are all of these different approaches.
We started applying this core dynamical system to these problems. And so the obvious thing to do, I think, was basically: you take that linear system—so this is the thing representing the temporal information—and then you put a non-linear layer afterwards, so you can then manipulate that temporal representation however you want. And you’ll learn that using normal backprop. And that’s what we call the Legendre Memory Unit.
More recently, people have taken that exact same structure and called it a state-space model, for obvious reasons—because basically, having a linear dynamical system and then a non-linear layer, that’s a state-space model. And that’s what BrainChip is using, for instance.
Also,
CE: And so for the last couple of years, that’s what we’ve really been focused on: building a chip that can natively run state-space models, run it extremely efficiently—because it’s specifically designed to do that—and fabricate that chip, and go to customers and start getting into all the many different products that you might be interested in having that in.
So that’s something that we’re really excited about, because we just got the chip back for it at the end of August. We got the chip back, we tested it, and we had it up and running and setting world records in low-power running state-space models within a week and a half.
SB: Talk more about this chip. I’m very interested!
CE: It’s not a neuromorphic chip, in the sense that it’s not event-based. It is a neural accelerator, so it has compute right near the memory. The memory is actually a little bit exotic—it’s called MRAM, so magnetoresistive RAM, which means that it’s non-volatile—so you can load your network on there, you can basically leave it, shut off, and it draws almost no power until you need it—and then it can run the model, which is really cool.
We’re able to do full speech recognition. So it could be sitting here basically typing out everything that I’m saying with about a hundred times less power than other edge hardware that’s available on the market—under 30 milliwatts. We can have it typing out whatever language you’re speaking in.
We can do other things with it, too. We can use it to control a device. So you can basically tell the device what you want to do, using natural language—you don’t have to memorize keywords or key phrases, you just say what you want. We’re also working with customers who want to use it to do things like monitor your biosignals—your heartbeat, your O2, your sweating, you know, anything. You can monitor all that and do on-device inference about the state of the person, warn them that they’re going to have a seizure if it’s EEG, or that they’re having some kind of heart palpitation, or what have you.
And we just started our early access program. So we’re working with customers, getting the hardware in their hands, helping them integrate that into their applications. We’re super excited about what this chip can do. It’s just kind of blowing the competition out of the water from a power and efficiency and performance perspective.
REC: An interesting aside, which is kind of like a reality check for me—a group of us went to speak to a DARPA director not long ago, and we were selling an idea of which one of the primary focuses was that it was basically extremely low power. And she could not care less about the power. She cared mostly about latency. That is the thing. And this was an embodied AI application, Giulia. She said, “I will burn all the power that I need to if you can get me the speed and the decisions to happen as quickly as possible, and as effectively as possible.” Which for me was like, “What!?” I expected that the power efficiency argument would’ve been a winner on the day.
It was interesting to me. We were trying to argue for small drones with small brains and so on, and she was like, eh, but what’s the latency? When do you make the decision? Anyway, very interesting.
A separate thing I don't recall people mentioning was the Youtube presentation he did a few weeks ago in which he went into his new chip in more detail. There were a few things worth noting such as listing some specs (eg. the chip uses 22nm node process, and a comparison with competitors:
View attachment 93607
A few takeaways:
- This gives further confirmation that ABR's LMU and Brainchip's State Space models are very similar competitive technologies.
- The rush to get Akida Cloud was likely partly in response to the threat ABR proposed. Akida's Gen 2 chip is still in production. However, while ABR have a chip now, Akida Gen 2 hardware was available to customers from 5th August (see link below), slightly earlier than ABR's chip. ABR got their chip back at the end of August, then had a few weeks of testing plus logistics before it could be provided to customers. Akida 2 would likely have been the FPGA prototype, so probably not the final version with maximum efficiency, but it allowed people to test at least a month earlier than ABR. In general, any new prototype chip can be released to customers this way, which provides an edge.
- In the Youtube clip, ABR were comparing their new chip to the 1st gen Akida and saying it wasn't comparable. It couldn't do more complex tasks like automatic speech recognition. I would guess Gen 2 woul have similar efficiency / performance to ABR's new chip unless there are significant improvements that Brainchip have made to hardware or algorithm efficiency (patent pending). Both chips are being manufactured in 22nm from memory so performance should be comparable when these values eventually become available.
- It's probably no surprise that ABR are working with customers to implement solutions. However, even if ABR have better hardware it doesn't mean customers will rush to adopt it over Akida Gen 2. Brainchip has spent many years building up their ecosystem and working with customers and has other advantages of it's own. Furthermore, Brainchip is more focused on IP than chips, which means they are working with some diffferent customers
- Brainchip's competitive advantage may not be as high as some people think. Akida Gen 3 will probably be important to ensuring they retain this
LAGUNA HILLS, Calif. – Aug 5th, 2025 – BrainChip Holdings Ltd (ASX: BRN, OTCQX: BRCHF, ADR: BCHPY), the world’s first commercial producer of ultra-low power, fully digital, event-based neuromorphic AI, today announced launch of the BrainChip Developer Akida Cloud, a new cloud-based access point to multiple generations and configurations of the company’s Akida™ neuromorphic technology. The initial Developer Cloud release will feature the latest version of Akida’s 2nd generation technology,BrainChip Launches Akida Cloud for Instant Neuromorphic Access
BrainChip launches Akida Cloud, providing instant access to the latest neuromorphic technology for developers and edge AI innovators.
| Date | Event | Capabilities / Significance | Source |
|---|---|---|---|
| 10 Sept 2018 | Announcement of Akida 1000 architecture | First neuromorphic SoC design unveiled; event‑based AI for video analytics, IoT, automotive | BrainChip Release |
| 2 Dec 2020 | ASX announcement: Completion of Akida production design | Final design ready for tape‑out; milestone toward commercial silicon | ASX release 2 Dec'20 |
| 21 Oct 2021 | ASX announcement: BrainChip begins taking orders for Akida AI Development Kits | Kits shipped to partners for evaluation and integration | ASX release. 21 Oct'21 |
| 9 Nov 2021 | BrainChip completes testing of production version of Akida chip | Validated performance and readiness for commercial deployment | ASX release. 9 Nov'21 |
| 28 Aug 2023 | First batch of Akida 1500 chips receivedfrom GlobalFoundries | Second silicon proof point; validated portability across foundries; built on 22nm FD‑SOI | BrainChip Release. See below. |
| 4 Nov 2025 | Launch of latest Akida 1500 Edge AI co‑processorat Embedded World North America | Achieves 800 GOPS under 300 mW; ideal for wearables, smart sensors, constrained environments | BusinessWire Release |
| Early 2026 (Roadmap) | Planned Akida Gen 2 tape‑out | Next‑generation neuromorphic IP; enhanced scalability, performance, and efficiency across automotive, IoT, wearables, defense | BrainChip roadmap |
Takeaway: The timeline from the 2018 architecture announcement → 2020 production design completion → Oct 2021 dev kit orders → Nov 2021 production testing completed→ 2023 Akida 1500 silicon → 2025 UPDATED AKIDA 1500 co‑processor launch → 2026 Gen 2 tape‑out milestone.A relevant interview regarding Chris Eliasmith of ABR, makes for an interesting read:
It has a few relevant parts such as:
SB: Your group has also come up with Legendre Memory Units to represent time. Listeners may remember that BrainChip have been using this approach as well. Can you talk about what they are and why they’re useful?
CE: So we were really working on, ‘How does the brain represent time?’ But we also came to realize that, well, this is a problem in machine learning. Time series is a massive area of research, and it’s really hard. And people have all kinds of different recurrent networks, such as LSTNs and GRUs and a million variants on each of these. Transformers are now what people are using, where it’s not a recurrent network but it kind of spreads time out in space so you can just process stuff in a feedforward way. So there are all of these different approaches.
We started applying this core dynamical system to these problems. And so the obvious thing to do, I think, was basically: you take that linear system—so this is the thing representing the temporal information—and then you put a non-linear layer afterwards, so you can then manipulate that temporal representation however you want. And you’ll learn that using normal backprop. And that’s what we call the Legendre Memory Unit.
More recently, people have taken that exact same structure and called it a state-space model, for obvious reasons—because basically, having a linear dynamical system and then a non-linear layer, that’s a state-space model. And that’s what BrainChip is using, for instance.
Also,
CE: And so for the last couple of years, that’s what we’ve really been focused on: building a chip that can natively run state-space models, run it extremely efficiently—because it’s specifically designed to do that—and fabricate that chip, and go to customers and start getting into all the many different products that you might be interested in having that in.
So that’s something that we’re really excited about, because we just got the chip back for it at the end of August. We got the chip back, we tested it, and we had it up and running and setting world records in low-power running state-space models within a week and a half.
SB: Talk more about this chip. I’m very interested!
CE: It’s not a neuromorphic chip, in the sense that it’s not event-based. It is a neural accelerator, so it has compute right near the memory. The memory is actually a little bit exotic—it’s called MRAM, so magnetoresistive RAM, which means that it’s non-volatile—so you can load your network on there, you can basically leave it, shut off, and it draws almost no power until you need it—and then it can run the model, which is really cool.
We’re able to do full speech recognition. So it could be sitting here basically typing out everything that I’m saying with about a hundred times less power than other edge hardware that’s available on the market—under 30 milliwatts. We can have it typing out whatever language you’re speaking in.
We can do other things with it, too. We can use it to control a device. So you can basically tell the device what you want to do, using natural language—you don’t have to memorize keywords or key phrases, you just say what you want. We’re also working with customers who want to use it to do things like monitor your biosignals—your heartbeat, your O2, your sweating, you know, anything. You can monitor all that and do on-device inference about the state of the person, warn them that they’re going to have a seizure if it’s EEG, or that they’re having some kind of heart palpitation, or what have you.
And we just started our early access program. So we’re working with customers, getting the hardware in their hands, helping them integrate that into their applications. We’re super excited about what this chip can do. It’s just kind of blowing the competition out of the water from a power and efficiency and performance perspective.
REC: An interesting aside, which is kind of like a reality check for me—a group of us went to speak to a DARPA director not long ago, and we were selling an idea of which one of the primary focuses was that it was basically extremely low power. And she could not care less about the power. She cared mostly about latency. That is the thing. And this was an embodied AI application, Giulia. She said, “I will burn all the power that I need to if you can get me the speed and the decisions to happen as quickly as possible, and as effectively as possible.” Which for me was like, “What!?” I expected that the power efficiency argument would’ve been a winner on the day.
It was interesting to me. We were trying to argue for small drones with small brains and so on, and she was like, eh, but what’s the latency? When do you make the decision? Anyway, very interesting.
A separate thing I don't recall people mentioning was the Youtube presentation he did a few weeks ago in which he went into his new chip in more detail. There were a few things worth noting such as listing some specs (eg. the chip uses 22nm node process, and a comparison with competitors:
View attachment 93607
A few takeaways:
- This gives further confirmation that ABR's LMU and Brainchip's State Space models are very similar competitive technologies.
- The rush to get Akida Cloud was likely partly in response to the threat ABR proposed. Akida's Gen 2 chip is still in production. However, while ABR have a chip now, Akida Gen 2 hardware was available to customers from 5th August (see link below), slightly earlier than ABR's chip. ABR got their chip back at the end of August, then had a few weeks of testing plus logistics before it could be provided to customers. Akida 2 would likely have been the FPGA prototype, so probably not the final version with maximum efficiency, but it allowed people to test at least a month earlier than ABR. In general, any new prototype chip can be released to customers this way, which provides an edge.
- In the Youtube clip, ABR were comparing their new chip to the 1st gen Akida and saying it wasn't comparable. It couldn't do more complex tasks like automatic speech recognition. I would guess Gen 2 woul have similar efficiency / performance to ABR's new chip unless there are significant improvements that Brainchip have made to hardware or algorithm efficiency (patent pending). Both chips are being manufactured in 22nm from memory so performance should be comparable when these values eventually become available.
- It's probably no surprise that ABR are working with customers to implement solutions. However, even if ABR have better hardware it doesn't mean customers will rush to adopt it over Akida Gen 2. Brainchip has spent many years building up their ecosystem and working with customers and has other advantages of it's own. Furthermore, Brainchip is more focused on IP than chips, which means they are working with some diffferent customers
- Brainchip's competitive advantage may not be as high as some people think. Akida Gen 3 will probably be important to ensuring they retain this
LAGUNA HILLS, Calif. – Aug 5th, 2025 – BrainChip Holdings Ltd (ASX: BRN, OTCQX: BRCHF, ADR: BCHPY), the world’s first commercial producer of ultra-low power, fully digital, event-based neuromorphic AI, today announced launch of the BrainChip Developer Akida Cloud, a new cloud-based access point to multiple generations and configurations of the company’s Akida™ neuromorphic technology. The initial Developer Cloud release will feature the latest version of Akida’s 2nd generation technology,BrainChip Launches Akida Cloud for Instant Neuromorphic Access
BrainChip launches Akida Cloud, providing instant access to the latest neuromorphic technology for developers and edge AI innovators.
You're right that I used a poor choice of terminology, I intended to refer to the roadmap early this year where AKD 2 ASIC tapeout was planned for January 2026.A few things.
Not surprising, there will be competitors, if there wasn't any, we should be very worried.
Quote from your post:
Akida's Gen 2 chip is still in production? No it's not, please elaborate what you mean? You mean development, right?
Akida Gen 2 hardware was available to customers from 5th August (see link below)? When it's in the cloud, it's not hardware, right?
Anyway, BRN have a much wider audience IMO.
A few snippets from GPT comparing the TSP1 with AKD1500 since they can be bought over the counter and not only IP (BRN):
When you'd pick TSP1 — when you'd pick AKD1500
- Choose TSP1 if your application is: voice / speech recognition, continuous audio processing, sensor-stream inference, always-on low-power audio/sensor device, simple embedded product — where you need a self-contained, low-power, reliable edge-AI SoC with deterministic latency.
- Choose AKD1500 if your application needs: high-performance edge AI, support for more complex / heavier models (vision, video, multi-sensor fusion), on-device learning, flexibility, future-proofing, or integration as a co-processor alongside a more capable host — especially when you want scalable performance and more computational headroom than a tiny embedded AI chip.
- TSP1 = “specialist, self-contained, efficient for time-series/audio/sensor embedded AI.”
- AKD1500 = “generalist, powerful, flexible, co-processor for demanding or varied edge AI tasks (vision, video, multiple modalities, ML-heavy workloads).”
TSP1 vs Akida1000 (processor)
When you’d want TSP1 vs when you’d want AKD1000
Use TSP1 if:
- Your application is primarily speech/audio processing, sensor data, continuous time-series, always-on voice UI, biosignals, low-power wearable devices, IoT — i.e. data is temporal, sequential, streaming.
- You need a self-contained, low-power SoC (MCU + NPU + memory + I/O) — minimal external dependencies, easy to embed in small form-factor devices.
- Power budget is tight (battery-powered, constrained devices), latency & real-time response matter, and the workload fits within modest neural model size.
Use AKD1000 if:
- You need flexibility and generality — want to run vision, audio + vision, sensor fusion, larger or more complex models than strictly small time-series nets.
- Your workload may benefit from event-based / sparsity-aware processing (e.g. event cameras, spatio-temporal sensor data, mixed modalities), or you anticipate growth/flexibility in model architecture over device lifetime.
- You are designing a system with a host processor (or can afford a co-processor architecture) and don’t need the chip to be a standalone SoC.
If I were to pick for voice-first, always-on edge device, TSP1 is the best fit.
- TSP1 = specialist, very efficient, minimal, optimized for time-series / streaming, low-power, embedded AI (speech, audio, sensors).
- AKD1000 = generalist neuromorphic/digital accelerator, more flexible and scalable — suited for vision, multimodal, heavier workloads, at the cost of greater system complexity and higher power when fully used.
If I were to build an edge AI box or sensor+vision device needing flexibility and higher compute, AKD1000 makes more sense.
Australia hasn't got a wicket all day, so could you please leave so we can get a wicket.Hope everyone is going to have a better day than me today as I have a ticket to watch the ashes
"Next year is our year" - every damn year for the last 8 yearsClearly 2026 is BrainChip year with Onsor and many other's bringing Akida to market.
Were going to need a bigger bank
Okay, no worries.
Honest truth I got bored and left and as soon as I left the gabba I heard an almighty roar lol
Totally agree, we need to get Gen 2 out there ASAP. The tape out of Gen 2 in early 2026 is important and hopefully it will allow customers who have FGPA to develop protypes in many instances.
www.meegle.com
fpgainsights.com
A relevant interview regarding Chris Eliasmith of ABR, makes for an interesting read:
It has a few relevant parts such as:
SB: Your group has also come up with Legendre Memory Units to represent time. Listeners may remember that BrainChip have been using this approach as well. Can you talk about what they are and why they’re useful?
CE: So we were really working on, ‘How does the brain represent time?’ But we also came to realize that, well, this is a problem in machine learning. Time series is a massive area of research, and it’s really hard. And people have all kinds of different recurrent networks, such as LSTNs and GRUs and a million variants on each of these. Transformers are now what people are using, where it’s not a recurrent network but it kind of spreads time out in space so you can just process stuff in a feedforward way. So there are all of these different approaches.
We started applying this core dynamical system to these problems. And so the obvious thing to do, I think, was basically: you take that linear system—so this is the thing representing the temporal information—and then you put a non-linear layer afterwards, so you can then manipulate that temporal representation however you want. And you’ll learn that using normal backprop. And that’s what we call the Legendre Memory Unit.
More recently, people have taken that exact same structure and called it a state-space model, for obvious reasons—because basically, having a linear dynamical system and then a non-linear layer, that’s a state-space model. And that’s what BrainChip is using, for instance.
Also,
CE: And so for the last couple of years, that’s what we’ve really been focused on: building a chip that can natively run state-space models, run it extremely efficiently—because it’s specifically designed to do that—and fabricate that chip, and go to customers and start getting into all the many different products that you might be interested in having that in.
So that’s something that we’re really excited about, because we just got the chip back for it at the end of August. We got the chip back, we tested it, and we had it up and running and setting world records in low-power running state-space models within a week and a half.
SB: Talk more about this chip. I’m very interested!
CE: It’s not a neuromorphic chip, in the sense that it’s not event-based. It is a neural accelerator, so it has compute right near the memory. The memory is actually a little bit exotic—it’s called MRAM, so magnetoresistive RAM, which means that it’s non-volatile—so you can load your network on there, you can basically leave it, shut off, and it draws almost no power until you need it—and then it can run the model, which is really cool.
We’re able to do full speech recognition. So it could be sitting here basically typing out everything that I’m saying with about a hundred times less power than other edge hardware that’s available on the market—under 30 milliwatts. We can have it typing out whatever language you’re speaking in.
We can do other things with it, too. We can use it to control a device. So you can basically tell the device what you want to do, using natural language—you don’t have to memorize keywords or key phrases, you just say what you want. We’re also working with customers who want to use it to do things like monitor your biosignals—your heartbeat, your O2, your sweating, you know, anything. You can monitor all that and do on-device inference about the state of the person, warn them that they’re going to have a seizure if it’s EEG, or that they’re having some kind of heart palpitation, or what have you.
And we just started our early access program. So we’re working with customers, getting the hardware in their hands, helping them integrate that into their applications. We’re super excited about what this chip can do. It’s just kind of blowing the competition out of the water from a power and efficiency and performance perspective.
REC: An interesting aside, which is kind of like a reality check for me—a group of us went to speak to a DARPA director not long ago, and we were selling an idea of which one of the primary focuses was that it was basically extremely low power. And she could not care less about the power. She cared mostly about latency. That is the thing. And this was an embodied AI application, Giulia. She said, “I will burn all the power that I need to if you can get me the speed and the decisions to happen as quickly as possible, and as effectively as possible.” Which for me was like, “What!?” I expected that the power efficiency argument would’ve been a winner on the day.
It was interesting to me. We were trying to argue for small drones with small brains and so on, and she was like, eh, but what’s the latency? When do you make the decision? Anyway, very interesting.
A separate thing I don't recall people mentioning was the Youtube presentation he did a few weeks ago in which he went into his new chip in more detail. There were a few things worth noting such as listing some specs (eg. the chip uses 22nm node process, and a comparison with competitors:
View attachment 93607
A few takeaways:
- This gives further confirmation that ABR's LMU and Brainchip's State Space models are very similar competitive technologies.
- The rush to get Akida Cloud was likely partly in response to the threat ABR proposed. Akida's Gen 2 chip is still in production. However, while ABR have a chip now, Akida Gen 2 hardware was available to customers from 5th August (see link below), slightly earlier than ABR's chip. ABR got their chip back at the end of August, then had a few weeks of testing plus logistics before it could be provided to customers. Akida 2 would likely have been the FPGA prototype, so probably not the final version with maximum efficiency, but it allowed people to test at least a month earlier than ABR. In general, any new prototype chip can be released to customers this way, which provides an edge.
- In the Youtube clip, ABR were comparing their new chip to the 1st gen Akida and saying it wasn't comparable. It couldn't do more complex tasks like automatic speech recognition. I would guess Gen 2 woul have similar efficiency / performance to ABR's new chip unless there are significant improvements that Brainchip have made to hardware or algorithm efficiency (patent pending). Both chips are being manufactured in 22nm from memory so performance should be comparable when these values eventually become available.
- It's probably no surprise that ABR are working with customers to implement solutions. However, even if ABR have better hardware it doesn't mean customers will rush to adopt it over Akida Gen 2. Brainchip has spent many years building up their ecosystem and working with customers and has other advantages of it's own. Furthermore, Brainchip is more focused on IP than chips, which means they are working with some diffferent customers
- Brainchip's competitive advantage may not be as high as some people think. Akida Gen 3 will probably be important to ensuring they retain this
LAGUNA HILLS, Calif. – Aug 5th, 2025 – BrainChip Holdings Ltd (ASX: BRN, OTCQX: BRCHF, ADR: BCHPY), the world’s first commercial producer of ultra-low power, fully digital, event-based neuromorphic AI, today announced launch of the BrainChip Developer Akida Cloud, a new cloud-based access point to multiple generations and configurations of the company’s Akida™ neuromorphic technology. The initial Developer Cloud release will feature the latest version of Akida’s 2nd generation technology,BrainChip Launches Akida Cloud for Instant Neuromorphic Access
BrainChip launches Akida Cloud, providing instant access to the latest neuromorphic technology for developers and edge AI innovators.
Haha. Same time as I posted. I only came on forum because cricket was a bit boring. Great minds hey. Between us we got 4 wickets
