BRN Discussion Ongoing

FJ-215

Regular
Hi Deng

This should not be seen as a problem. The failure to publish is a product of a highly ethical company building trusted relationships with its customers and the scientific community generally.

The papers that have been published in respect of AERO using the ADE to identify the 20 gas data set WITH A STATE OF THE ART PERFORMANCE has never been questioned.

I can guarantee that academics and competitors would have been secretly running the experiment again and again to find a flaw so they could publish that the science was flawed. It was published in 2019 and not a single negative response.

The same thing would be taking place with the more recent paper regarding Malt. To date nothing negative.

Peer reviewed scientific publications are incredibly important and cannot be fudged in anyway otherwise the scientists involved can have their reputations and credibility destroyed.

Brainchip is releasing information all the time. When Peter van der Made presented in November 2020 at BIC that AKD1000 engineering sample was 178 times more power efficient than Loihi that was a very brave statement to make.

It is now 2022 and there has never been any question from Intel that this statement was wrong or flawed in some fashion.

The science is there for the market to see and the validation is also there with the companies who are early adopters Mercedes, Valeo, NASA, Vorago and licensees Renesas and MegaChips.

The share price is what it is and is nothing to do with anything at the moment other than world wide economic and political tensions and some activist shorting.

My opinion only DYOR
FF

AKIDA BALLISTA
Hi FF,

Peer reviewed scientific publications are incredibly important and cannot be fudged in anyway otherwise the scientists involved can have their reputations and credibility destroyed.

A few years ago I would agreed with this but sadly I no longer believe it to be true. Ego, money & religion have combined to pollute true science unfortunately. Tobacco, oil, food. If there is money to be made, junk science is a powerful tool.

Have been down the rabbit hole of the diet world the past year. Been an eye opener.

Sugar vs. Cholesterol: John Yudkin vs. Ancel Keys

To relate this back to BRN. Everywhere the Western Diet goes, poor health seems to follow. I see smart health as potentially our biggest market.
 
  • Like
Reactions: 13 users
Hi tls
The best person to speak to this would be Peter van der Made however we have the second best in Dio so I will leave it to him to deal with in detail. He has addressed this issue before and has kept the slide put out by Brainchip with the plan for these devices.

I can say in a purely non technical sense that AKD500 was a cheaper smaller version of AKD1000 aimed at white goods like fridges. AKD1500 quoting Dio was AKD1000 with a little bit of LSTM added. AKD2000 has the full blown LSTM on board a must for ADAS and AV allowing for prediction from given events ie; ball rolling out from between parked cars could be followed by child, stroller coming out from parked cars, drunk staggering around at edge of the road could fall into path of vehicle type of things.

AKD3000 is the biggie as it is looking to implement a cortical column for advanced thinking on chip so that the current cochlear implant for example could produce real sounds for the wearer or it could be used as the foundation for a bionic eye that restores vision. I believe AKD4000 and AKD5000 are approaching true artificial general intelligence and will replace Fact Finder on here.

Seriously though Dio is the best one to give you the story.

My opinion only DYOR
FF

AKIDA BALLISTA

Denzel Washington GIF
 
  • Like
Reactions: 4 users

JK200SX

Regular
Hi tls
The best person to speak to this would be Peter van der Made however we have the second best in Dio so I will leave it to him to deal with in detail. He has addressed this issue before and has kept the slide put out by Brainchip with the plan for these devices.

I can say in a purely non technical sense that AKD500 was a cheaper smaller version of AKD1000 aimed at white goods like fridges. AKD1500 quoting Dio was AKD1000 with a little bit of LSTM added. AKD2000 has the full blown LSTM on board a must for ADAS and AV allowing for prediction from given events ie; ball rolling out from between parked cars could be followed by child, stroller coming out from parked cars, drunk staggering around at edge of the road could fall into path of vehicle type of things.

AKD3000 is the biggie as it is looking to implement a cortical column for advanced thinking on chip so that the current cochlear implant for example could produce real sounds for the wearer or it could be used as the foundation for a bionic eye that restores vision. I believe AKD4000 and AKD5000 are approaching true artificial general intelligence and will replace Fact Finder on here.

Seriously though Dio is the best one to give you the story.

My opinion only DYOR
FF

AKIDA BALLISTA

"I believe AKD4000 and AKD5000 are approaching true artificial general intelligence and will replace Fact Finder on here."

Hmmm, I don't think you'll get replaced. They'd need an AKD88000 or better :)
 
  • Like
  • Haha
  • Fire
Reactions: 26 users

Diogenese

Top 20
Is GrAI Matter Labs a competitor Chip?
Hi Dang Son,

We have looked at GraiMatter a few times before, and yes, they are a competitor.

As a result of your query I revisited this paper:

Spiking Neural Networks: Research Projects Or Commercial Products?​

742Shares
sharethis sharing button

Opinions differ widely, but in this space that isn’t unusual.
MAY 18TH, 2020 - BY: BRYON MOYER
https://semiengineering.com/spiking-neural-networks-research-projects-or-commercial-products/

which includes a conversational comparison of Akida and GraiMatter (GM).

Some points of differentiation:

Akida uses rank coding as explained in the PCT patent application WO20210027152,

1645924238316.png


whereas GM encodes information in a 16-bit byte:
All of these coding approaches aside, GrAI Matter uses a more direct approach. “We encode values directly as numbers – 8- or 16-bit integers in GrAI One or Bfloat16 in our upcoming chip. This is a key departure from other neuromorphic architectures, which have to use rate or population or time or ensemble codes. We can use those, too, but they are not efficient,” said Tapson.

The need to perform 16-bit maths may require 264 operations per byte multiplication, whereas, Akida in 4-bit mode would only need to perform 16 operations, giving Akida a very large advantage in both speed and power consumption. Akida in 1-bit mode has a still greater advantage, requiring only a single operation.

Akida does one-shot, on-chip learning.

GrAI Matter also is doing an all-digital implementation. It uses an on-chip packet-switched network to route the “spikes.” GrAI Matter’s overall architecture is shown below (the node implementation is shown above in Figure 5). The company trains its chip using classical techniques, converting the result to the GrAL Matter format for implementation.

One other major difference is that Akida and event-based cameras (DVS) are a match made in silicon heaven.

Even though this [# GraiMatter #] is an event-based engine, the network has been optimized to deal with standard video streams instead of DVS event streams. In a manner similar to the ISSCC paper discussed in a prior article, these operate on the differences between frames rather than the full frames. That “diff” is taken both at the input and at each activation layer, creating an enormous amount of sparsity entering and flowing through the network.

Fig. 8: GrAI Matter processes only changed pixels in each successive layer. Source: GrAI Matter Lab
s


This means that GraiMatter must do a comparison with all the pixels in successive frames to identify which pixels have changed (an event), whereas, in the case of DVS, the DVS sensor automatically presents "events" at its output.
 
Last edited:
  • Like
  • Fire
Reactions: 38 users

Diogenese

Top 20
Hi tls
The best person to speak to this would be Peter van der Made however we have the second best in Dio so I will leave it to him to deal with in detail. He has addressed this issue before and has kept the slide put out by Brainchip with the plan for these devices.

I can say in a purely non technical sense that AKD500 was a cheaper smaller version of AKD1000 aimed at white goods like fridges. AKD1500 quoting Dio was AKD1000 with a little bit of LSTM added. AKD2000 has the full blown LSTM on board a must for ADAS and AV allowing for prediction from given events ie; ball rolling out from between parked cars could be followed by child, stroller coming out from parked cars, drunk staggering around at edge of the road could fall into path of vehicle type of things.

AKD3000 is the biggie as it is looking to implement a cortical column for advanced thinking on chip so that the current cochlear implant for example could produce real sounds for the wearer or it could be used as the foundation for a bionic eye that restores vision. I believe AKD4000 and AKD5000 are approaching true artificial general intelligence and will replace Fact Finder on here.

Seriously though Dio is the best one to give you the story.

My opinion only DYOR
FF

AKIDA BALLISTA
Great summary - I rest on my briefs.
 
  • Haha
  • Like
  • Wow
Reactions: 18 users
  • Like
  • Love
Reactions: 19 users

Diogenese

Top 20
  • Haha
  • Like
Reactions: 14 users

Diogenese

Top 20
Last edited:
  • Like
  • Haha
Reactions: 12 users

Dang Son

Regular
Every producer of semiconductors is a competitor to Brainchip.

The better question to ask is can GrAi compete with Brainchip. The answer to which is no not at present.

GrAi are doing a simple SNN chip. Brainchip is doing an SCNN chip. Brainchip does one shot and incremental learning and CNN to SNN conversion on chip. AKIDA can as a result learn and adapt and be taught new information after commencing operation.

On GrAi's site it refers to the advantages of using SNN in a camera door bell from a power perspective. GrAi using SNN saves power on your door bell by only processing relevant events and is told at the factory what the relevant events are that it needs to process. Great idea. Saves power.

Brainchip's AKIDA on the other hand uses SNN and saves power however as Brainchip's Anil Mankar and others have stated it can be shown the images (or the real persons) who live at the address or who are permitted access and it will then only react to those persons who are not entitled to be there and ring the bell. If after some months Christmas arrives and guests come to stay you can add their images for the duration of their stay. If you sell the property the system might be considered a fixture and need to be left. You could cancel all the existing images and the new owners can then add themselves.

On top of this GrAi appears to only be capable of processing one sense at a time whereas Brainchip makes the point that it can process all five senses plus Lidar, Radar, Sonar etc on AKD1000.

So as I said at the beginning the question is not are they a competitor but can they compete.

The other thing to consider is something which over on HC the trolls used to attack Brainchip on was the low cost of the AKD1000 at $US10 to $US15. The thing about this is that by being so cheap and even much, much cheaper if you take only a couple of nodes via the IP sales model you can get all these features for a ridiculously low price so that even if a company like GrAi is tendering along side Brainchip their more powerful self learning one shot training chip will be the better choice.

If you go into a car dealer and they say this is the base model but you can have the AMG version with all the upgrades for exactly the same price I would suggest on 99 out of 100 occasions you are going to take the AMG version.

My opinion only DYOR
FF

AKIDA BALLISTA
Thanks FF , a terrific comparison/explanation
I'll take the AMG, while it's at a ridiculously low price and hopefully it has the engine signed by its maker.
 
  • Like
  • Haha
Reactions: 5 users
This Akida development timetable is from mid-2020:
View attachment 1885

Did they hit the 2020 and 2021 targets; and if so are these included in the existing EAPs/NDAs? Apologies if this is common knowledge and I have simply missed it
 
  • Like
Reactions: 2 users

Dang Son

Regular
Hi Dan Son,

We have looked at GraiMatter a few times before, and yes, they are a competitor.

As a result of your query I revisited this paper:

Spiking Neural Networks: Research Projects Or Commercial Products?​

742Shares
sharethis sharing button

Opinions differ widely, but in this space that isn’t unusual.
MAY 18TH, 2020 - BY: BRYON MOYER
https://semiengineering.com/spiking-neural-networks-research-projects-or-commercial-products/

which includes a conversational comparison of Akida and GraiMatter (GM).

Some points of differentiation:

Akida uses rank coding as explained in the PCT patent application WO20210027152,

View attachment 1882

whereas GM encodes information in a 16-bit byte:
All of these coding approaches aside, GrAI Matter uses a more direct approach. “We encode values directly as numbers – 8- or 16-bit integers in GrAI One or Bfloat16 in our upcoming chip. This is a key departure from other neuromorphic architectures, which have to use rate or population or time or ensemble codes. We can use those, too, but they are not efficient,” said Tapson.

The need to perform 16-bit maths may require 264 operations per byte multiplication, whereas, Akida in 4-bit mode would only need to perform 16 operations, giving Akida a very large advantage in both speed and power consumption. Akida in 1-bit mode has a still greater advantage, requiring only a single operation.

Akida does one-shot, on-chip learning.

GrAI Matter also is doing an all-digital implementation. It uses an on-chip packet-switched network to route the “spikes.” GrAI Matter’s overall architecture is shown below (the node implementation is shown above in Figure 5). The company trains its chip using classical techniques, converting the result to the GrAL Matter format for implementation.

One other major difference is that Akida and event-based cameras (DVS) are a match made in silicon heaven.

Even though this [# GraiMatter #] is an event-based engine, the network has been optimized to deal with standard video streams instead of DVS event streams. In a manner similar to the ISSCC paper discussed in a prior article, these operate on the differences between frames rather than the full frames. That “diff” is taken both at the input and at each activation layer, creating an enormous amount of sparsity entering and flowing through the network.

Fig. 8: GrAI Matter processes only changed pixels in each successive layer. Source: GrAI Matter Lab
s


This means that GraiMatter must do a comparison with all the pixels in successive frames to identify which pixels have changed (an event), whereas, in the case of DVS, the DVS sensor automatically presents "events" at its output.
Thanks heaps Dio, for going to all that time and effort.
You're a Genius researcher and such a valuable contributor to our community.
My take is GrAi Matter falls way short of AKIDA capabilities,
Has to be pretrained
Heavy on network reliance
Heavy on data transfer
and likely very heavy on power consumption, comparatively
So, no competition for AKIDA
Or as FF would put it, GrAI Matter can't compete with AKIDA.
Yee-Ha !!!
 
  • Like
Reactions: 13 users

Diogenese

Top 20
Did they hit the 2020 and 2021 targets; and if so are these included in the existing EAPs/NDAs? Apologies if this is common knowledge and I have simply missed it
The hardware development has had a couple of delays, the most significant being the redesign of Akida 1000 to include new features such as CNN2SNN and optional 2-bit and 4-bit weights and activations. These were made to take into account the feedback after the engineering version of Akida 1000 SoC was sent to EAP clients.

On the other hand, @Fact Finder reminds us that Akida 2000 is being breadboarded in a benchtop model, nd work has begun on Akida 3000 with the cortical network.

https://stocksdownunder.com/video/b...-semiconductor-conference-on-1-november-2021/

20211130
Marc: Okay. Regarding the Akida 2000 and the timeline specifically for that one, can you talk a little bit about the development timeline?

Marc: Yes. Akida 2000 is very exciting too. We are developing that here in Perth. We have an excellent team here in Perth of highly qualified people who are working on Akida 2000 and Akida 3000. So we’re really thinking ahead here. Akida 2000 will be optimized for more complex network architectures, such as [inaudible 00:29:37]
{#LSTM#}* transformers. And for a very large part, those simulations of those chips are… At least Akida 2000 is already…the simulation is working already. And we are getting ready to hand it over to engineering. So Akida 3000 is going to a cortical network. Cortical networks are based on the way the human cortex works. They’re still a lot of open questions about, especially in neuroscience, how that exactly works. We’ll be building models of the cortex and see how we can apply that to a real commercial environment. Of course, everything we do, of course, even though the science is fascinating, we have to always keep in mind that we are building commercial products.

Marc: All right. One interesting question I saw earlier coming from Nathan about the advantage you have in terms of time versus competitors. Previously you stated two to three years advantage versus competitors. With the Akida 1000 now almost sort of commercially available, do you still think your timeline or the time advantage is still two to three years or has that changed?

Peter: Yes, we are working very hard to make sure that we maintain that advantage. So Akida 1000, we estimate to be two or three years ahead of the market. For instance, we have convolution and hardware not by a processor. We have real-time learning that nobody else has been able to manage. So we definitely have a two or three advantage in Akida 1000, plus we have commercializing that product. But to maintain that advantage, we are working on Akida 2000 and 3000. So we stay ahead of the markets. With Akida 3000, we’re probably about five or six years ahead at the moment
.

* I have deciphered the inaudible words as "LSTM" - Long Short Term Memory which is used to identify movement of objects in the field of view. Events from the previous view are compared with events from the current view to detect movement. This necessitates temporarily storing the previous events until the comparison has taken place. LSTM is critical for AV and ADAS.
 
Last edited:
  • Like
  • Fire
  • Love
Reactions: 27 users

FJ-215

Regular
Hi FF,

Peer reviewed scientific publications are incredibly important and cannot be fudged in anyway otherwise the scientists involved can have their reputations and credibility destroyed.

A few years ago I would agreed with this but sadly I no longer believe it to be true. Ego, money & religion have combined to pollute true science unfortunately. Tobacco, oil, food. If there is money to be made, junk science is a powerful tool.

Have been down the rabbit hole of the diet world the past year. Been an eye opener.

Sugar vs. Cholesterol: John Yudkin vs. Ancel Keys

To relate this back to BRN. Everywhere the Western Diet goes, poor health seems to follow. I see smart health as potentially our biggest market.

Hopefully I'm not some sort of tin foil hat clad loony. Then again how would I know???
This quote is from John Nash (A Beautiful Mind) about his mental illness and how a genius could believe in aliens.

“Because the ideas I had about supernatural beings came to me the same way that my mathematical ideas did,” replied Nash. “So I took them seriously.”

Anyways, I'm no genius but having found BRN I hope my brain is capable of a reasonable degree of logic.

Here is a bit of an insight as to why I think smart health will be such a big market for BRN. There are members of the medical community that are now gunning for the sugar industry like they did for tobacco. How long did it take to confirm that smoking causes cancer? Is sugar worse than tobacco? This video is over an hour long but just pick up the 5 minutes or so from the 16.50 minute mark.



Ps. Berry pushes a carnivore diet, Lustig is just anti sugar (re: fructose) .
 
Last edited:
  • Like
Reactions: 4 users
Hi FF,

Peer reviewed scientific publications are incredibly important and cannot be fudged in anyway otherwise the scientists involved can have their reputations and credibility destroyed.

A few years ago I would agreed with this but sadly I no longer believe it to be true. Ego, money & religion have combined to pollute true science unfortunately. Tobacco, oil, food. If there is money to be made, junk science is a powerful tool.

Have been down the rabbit hole of the diet world the past year. Been an eye opener.

Sugar vs. Cholesterol: John Yudkin vs. Ancel Keys

To relate this back to BRN. Everywhere the Western Diet goes, poor health seems to follow. I see smart health as potentially our biggest market.
Agree there is junk science but a little investigation can usually reveal it to be just that.

In Brainchip’s case however having competitors like Intel with all the reasons in the world to expose flaws in Brainchip’s science I have absolute confidence in the legitimacy of its published claims.

My opinion only of course so DYOR
FF

AKIDA BALLISTA
 
  • Like
  • Love
Reactions: 15 users

Dang Son

Regular
Hopefully I'm not some sort of tin foil hat clad loony. Then again how would I know???
This quote is from John Nash (A Beautiful Mind) about his mental illness and how a genius could believe in aliens.

“Because the ideas I had about supernatural beings came to me the same way that my mathematical ideas did,” replied Nash. “So I took them seriously.”

Anyways, I'm no genius but having found BRN I hope my brain is capable of a reasonable degree of logic.

Here is a bit of an insight as to why I think smart health will be such a big market for BRN. There are members of the medical community that are now gunning for the sugar industry like they did for tobacco. How long did it take to confirm that smoking causes cancer? Is sugar worse than tobacco? This video is over an hour long but just pick up the 5 minutes or so from the 16.50 minute mark.



Ps. Berry pushes a carnivore diet, Lustig is just anti sugar.

AKIDA combined with sensor in the toilet bowl ,
for full analysis of a user's waste products.😃
The Smart Toilet, interphased by the individual's Bluetooth App
 
  • Like
  • Fire
Reactions: 6 users

FJ-215

Regular
Agree there is junk science but a little investigation can usually reveal it to be just that.

In Brainchip’s case however having competitors like Intel with all the reasons in the world to expose flaws in Brainchip’s science I have absolute confidence in the legitimacy of its published claims.

My opinion only of course so DYOR
FF

AKIDA BALLISTA
Agree there is junk science but a little investigation can usually reveal it to be just that.

Yes.

Need to look at the time scales though, money to be made while the arguments are being resolved. As a lawyer did you ever employ similar tactics?
 

FJ-215

Regular
AKIDA combined with sensor in the toilet bowl ,
for full analysis of a user's waste products.😃
The Smart Toilet
Some things I'm not looking forward to in the future.

Bad enough now getting a lecture from my GP every 6 months, don't need it every morning from the crapper!!
 
  • Like
  • Haha
Reactions: 8 users

Terroni2105

Founding Member

1645933333141.jpeg


1645933375709.jpeg
 
  • Like
  • Fire
  • Love
Reactions: 42 users

M_C

Founding Member

Attachments

  • Screenshot_20220227-134727_LinkedIn.jpg
    Screenshot_20220227-134727_LinkedIn.jpg
    930.9 KB · Views: 82
  • Like
  • Fire
Reactions: 22 users

Bravo

If ARM was an arm, BRN would be its biceps💪!
Hey gang!

In this article dated 25 Feb 2022 it says "Valeo is going all-in on a business strategy focused on electrification and advanced driver-assistance systems to secure its place in a market it believes will see an “explosion” in growth.

Explosion in growth? Now where have I heard that expression used before? 😝

Good times!


(Extract)

“Our strategy is very simple and straight forward. Mobility will become carbon-free and will become safer,” Périllat says. “The technologies behind this are going to be electrification and ADAS or driving assistance. The growth and use of these technologies will, literally, explode in the coming years.”

He says we are only at the beginning of a period of what he calls hypergrowth for these two pivotal markets for the automotive industry.

“I think the growth in electrification and ADAS will last one or two decades and be a growth that we have never seen before,” Périllat continues. “In the next 15 years the electrification market will reach €200 billion ($224 billion), while the ADAS market will reach €120 billion ($134 billion).


 
  • Like
  • Fire
  • Love
Reactions: 35 users
Top Bottom