BRN Discussion Ongoing

[DingoBorat]

So they said they can do transformers as well on their SNN.
As per brainchip we can also take transformer load with akida 2000 but the same is do able because of TENNs.
On top brainchip also said earlier the results are very encouraging when comparing akida with gpt 2.
Is there a co incidence??
On top the processor used is samsung on 28 nm.

I don't think Kaist, has anything to do with us, personally.

Process size, doesn't mean anything, it's just a good proven one, that doesn't cost as much as the smaller ones (nobody, is going to produce "research chips" in 7nm for example).

I don't think their chip is pure digital either (I think @Diogenesed looked into it?)..

No surprise, that Samsung is involved and I believe they have a history with us, but that is one of their main foundries.

Any input from BrainChip, is inspired in my opinion.

And I'd Love for Samsung, to be onboard.

 

[Diogenese]

I don't think Kaist, has anything to do with us, personally.

Process size, doesn't mean anything, it's just a good proven one, that doesn't cost as much as the smaller ones (nobody, is going to produce "research chips" in 7nm for example).

I don't think their chip is pure digital either (I think @Diogenesed looked into it?)..

No surprise, that Samsung is involved and I believe they have a history with us, but that is one of their main foundries.

Any input from BrainChip, is inspired in my opinion.

And I'd Love for Samsung, to be onboard.

Yes. KAIST are into analog. The term "in-memory compute" is usually used in relation to analog, in that the calculations are performed by the memory circuits by accumulating a voltage whose amplitude is proportional to the number of input signals.

 

[KKFoo]

Because of Akida's 4-bit capability, it can have "home-made" models which are more compact than the standard 8-bit models, and, of course, Akida can handle the 8-bit models as well, with the accompanying reduction in efficiency advantage. And let's not forget Akida's 1 and 2-bit capability for super low power consumption.

I think that developing Akida-specific models for Valeo, Mercedes and others is where a lot of our effort will be focussed. These will not be universal models. They will be more functionally biassed.

Hi Diogenese, I believe you are best person for this question.. Are the VVDN edge box use for research and development purposes or it can be used by end user? Let's say I want to set up a face recognition security system in my office, can I just buy an edge box and plug it into my camera system and it is ready to go or I still need to go and develop some software system to interact with the edge box?
Thank you in advance if you can provide me with the answer..

 

[Diogenese]

Hi Diogenese, I believe you are best person for this question.. Are the VVDN edge box use for research and development purposes or it can be used by end user? Let's say I want to set up a face recognition security system in my office, can I just buy an edge box and plug it into my camera system and it is ready to go or I still need to go and develop some software system to interact with the edge box?
Thank you in advance if you can provide me with the answer..

Hi KK,

The Edge Boxes are definitely suitable for end use. Of course they can also be used for R&D, but they are the real thing.

They come with pre-made models, but you can also develop your own model library or adapt the pre-made ones using on-chip learning.

 

[rgupta]

I don't think Kaist, has anything to do with us, personally.

Process size, doesn't mean anything, it's just a good proven one, that doesn't cost as much as the smaller ones (nobody, is going to produce "research chips" in 7nm for example).

I don't think their chip is pure digital either (I think @Diogenesed looked into it?)..

No surprise, that Samsung is involved and I believe they have a history with us, but that is one of their main foundries.

Any input from BrainChip, is inspired in my opinion.

And I'd Love for Samsung, to be onboard.

The reason I try to relate is not only 28nm chip, but also a report by brainchip where they compare akida with chat gpt2. But the biggest shock to me is they said their chip can do transformers though 1000 was not able to perform that function but akida 2 can do the same.
Anyway at the end a lot of wires are entangled and only time will have all the answers. But one thing for sure there is a lot of happening behind the scenes and still no concrete news.

 

[DingoBorat]

The reason I try to relate is not only 28nm chip, but also a report by brainchip where they compare akida with chat gpt2. But the biggest shock to me is they said their chip can do transformers though 1000 was not able to perform that function but akida 2 can do the same.
Anyway at the end a lot of wires are entangled and only time will have all the answers. But one thing for sure there is a lot of happening behind the scenes and still no concrete news.

It definitely is, a tangled web out there Rgupta..

That's what makes the dot joining Fun, Frustrating and potentially Fruitless, all at the same time..

 

[DingoBorat]

유회준 교수 연구팀, 뉴로모픽 컴퓨팅 기반의 초저전력 거대 언어 모델 인공지능반도체 핵심기술 ‘상보형-트랜스포머’ 개발 | KAIST 반도체시스템공학과

KAIST 반도체시스템공학과

sse.kaist.ac.kr

Used Google translator from Korean to English
Their SNN also compairing on GPT-2

Through this, it was possible to reduce the parameters of the GPT-2 giant model from 708 million to 191 million and the parameters of the T5 model used for translation from 402 million to 76 million. As a result of this compression work, we succeeded in reducing the power consumed by loading language model parameters from external memory by 70%. According to the researchers, the complementary transformer consumes 1/625th the power of the NVIDIA A100 GPU, while enabling high-speed operation of 0.4 seconds for language generation using the GPT-2 model and 0.2 seconds for language translation using the T5 model. In the case of language generation due to parameter lightweighting, the accuracy decreased by 1.2 branching coefficient (lower means that the language model was learned better), but the researchers explained that it is at a level where people will not feel awkward when reading the generated sentences. In the future, the research team plans to expand the scope of neuromorphic computing to various application fields rather than limiting it to language models.

Sounds like a bit of stretch now, saying it's running ChatGPT2, when it's actually a compressed model of it, with just over a quarter of the parameters...

Would be more accurate I think, to say it was running a SLM "derived" from ChatGPT2..

 

[rgupta]

It definitely is, a tangled web out there Rgupta..

That's what makes the dot joining Fun, Frustrating and potentially Fruitless, all at the same time..

That is a screen shot of Dr Tony Lewis at AGM where akida was compared with Chat gpt 2 and it was claimed our model can use 5000 times less power hungry and can run on the edge.
And then team brainchip raised last CR just to work on LLMs for neurophonic chips.
I donot know why everyone is comparing with chat gpt 2 while we already have chat gpt 4 plus.
But again for joining but yes Dr Tony Lewis was saying almost the same thing.

 

[DingoBorat]

That is a screen shot of Dr Tony Lewis at AGM where akida was compared with Chat gpt 2 and it was claimed our model can use 5000 times less power hungry and can run on the edge.
And then team brainchip raised last CR just to work on LLMs for neurophonic chips.
I donot know why everyone is comparing with chat gpt 2 while we already have chat gpt 4 plus.
But again for joining but yes Dr Tony Lewis was saying almost the same thing.

It's to do with the number of parameters Rgupta.

ChatGPT4 is huge..

Size Matters: How Big Is Too Big for An LLM?

Compute-optimal large language models according to the Chinchilla paper

pub.towardsai.net

I was always told size didn't matter ..

No chance at all, of running anything like ChatGPT4, outside of a data center, with current technology.

Interestingly, this shows over twice the parameters, for the full ChatGPT2, that Kaist claims it has (and they are running with just over a quarter of that).

 

[stockduck]

https://www.notebookcheck.net/Dell-Latitude-5455-arrives-later-this-month-in-Snapdragon-X-Plus-X1P-64-100-and-X1P-42-100-options.883656.0.html

"....
Being a business-centric device, Dell offers several security features on the Latitude 5455 including SafeBIOS, IR biometric login, optional fingerprint reader in the power button, camera privacy shutter, and optional security hardware authentication bundles.

Customers can also configure additional hardware security measures including a chassis intrusion switch, hard drive wipes, and tamper-evident packaging along with software measures such as Crowdstrike (ahem!), Secureworks, and Netskope.

The Latitude 5455 uses Qualcomm's FastConnect 7800 Wi-Fi 7 WLAN card with Bluetooth 5.4 for fast wireless networking. Additionally, users also get 2x USB4 Type-C 40 Gbps ports, 1x USB 3.2 Gen1 Type-A with Power Share, a microSD card reader, and a combo audio jack."


Well,..... that sounds interesting to me...what the hell is Dell doing?
But there is no specific point for Brainchip IP insn`t it?

Here a description in german language....:

Dell Latitude 5455: Business-Arbeiter startet mit Qualcomm-Rechenchip und Copilot

Dell bringt ein neues Notebook mit einem Rechenchip von Qualcomm auf den Markt. Mit diesen dürfte sich das Unternehmen an Kunden richten, die ein Notebook für eher allgemeine Aufgaben suchen und etwa von einer langen Akkulaufzeit profitieren möchten.

www.notebookcheck.com

"...Dell is launching a new notebook with a computing chip from Qualcomm. With these, the company is likely to be aimed at customers who are looking for a notebook for more general tasks and who would like to benefit from a long battery life.
....
Users should benefit from various options, such as a search function with input in natural language or a special suppression of disturbing ambient noise.
....."

translated from google translator

 

[rgupta]

It's to do with the number of parameters Rgupta.

ChatGPT4 is huge..

Size Matters: How Big Is Too Big for An LLM?

Compute-optimal large language models according to the Chinchilla paper

pub.towardsai.net

I was always told size didn't matter ..

View attachment 68940


No chance at all, of running anything like ChatGPT4, outside of a data center, with current technology.

Interestingly, this shows over twice the parameters, for the full ChatGPT2, that Kaist claims it has (and they are running with just over a quarter of that).

There is no doubt it is almost impossible to make a 1trillion parter model on the edge. But do we need all trillion parameters all the time. It is just like I have 10 wardrobes full of clothes but I can only wear a few at a given time. Which arises the question of what is important in a given situation, in a given industry etc.
There are a lot of instances where we may not need full chat gpt and that is why it is expected industry specific Small language models will be the future.

 

[Frangipani]

Boost Your Radar-Lidar Systems with Akida | BrainChip posted on the topic | LinkedIn

Optimize Your Radar-Lidar Systems with BrainChip’s Event-based Akida for AI Computing Developing radar-lidar solutions that demand low latency and high accuracy? BrainChip’s Akida event-based compute platform is designed for early detection and improved tracking while featuring efficient...

www.linkedin.com

Radar Lidar Applications Whitepaper

The Advantages of Event-Based Computing for Radar/LiDAR Applications In the rapidly evolving field of radar technology, the demand for more efficient, responsive, and adaptable systems is ever-growing. Traditional radar systems have served us well for decades, but the advent of event-based...

brainchip.com

 

[supersonic001]

Boost Your Radar-Lidar Systems with Akida | BrainChip posted on the topic | LinkedIn

Optimize Your Radar-Lidar Systems with BrainChip’s Event-based Akida for AI Computing Developing radar-lidar solutions that demand low latency and high accuracy? BrainChip’s Akida event-based compute platform is designed for early detection and improved tracking while featuring efficient...

www.linkedin.com

 

[Cgc516]

The best we can do to let the shorters down to the hell!

 

[keyeat]

It's to do with the number of parameters Rgupta.

ChatGPT4 is huge..

Size Matters: How Big Is Too Big for An LLM?

Compute-optimal large language models according to the Chinchilla paper

pub.towardsai.net

I was always told size didn't matter ..

View attachment 68940


No chance at all, of running anything like ChatGPT4, outside of a data center, with current technology.

Interestingly, this shows over twice the parameters, for the full ChatGPT2, that Kaist claims it has (and they are running with just over a quarter of that).

Bigger isn’t always better: How hybrid AI pattern enables smaller language models | IBM

Bigger is not always better, as specialized models outperform general-purpose models with lower infrastructure requirements.

www.ibm.com

"

Bigger isn’t always better: How hybrid AI pattern enables smaller language models​

"

Why need a model to tell you about Shakespeare when its main function is to summarise input data / find trends / forecast ....

 

[Labsy]

What’s your final amount you are going for labsy?

Hey man, I'm just gobbling up small parcels sub 16c. I don't really have an end goal... Haven't sold a share since covid. I suspect insto traders are passing to and fro large bundles and making penies on the dollar with their algos. Possibly some retail traders also trading large amounts daily too. I'm just picking up when I can. Just in top 100 - 150 ATM... Comfortable here but actively accumulating so who knows brother.

 

[Shadow59]

Hey man, I'm just gobbling up small parcels sub 16c. I don't really have an end goal... Haven't sold a share since covid. I suspect insto traders are passing to and fro large bundles and making penies on the dollar with their algos. Possibly some retail traders also trading large amounts daily too. I'm just picking up when I can. Just in top 100 - 150 ATM... Comfortable here but actively accumulating so who knows brother.

I've been slowly accumulating too. Roughly what number of shares would be the top 150 or 100. I now have way more than I originally intended.

 

[wilzy123]

I donot know why everyone is comparing with chat gpt 2 while we already have chat gpt 4 plus

Completely different models... completely different power/resource requirements... completely different applications. Brainchip are simply demonstrating Akida's capability and strengths in the large language model (language based applications) space, using GPT2... a small language model as an example. It's not reasonable to compare GPT2/4 simply on the basis of how recently they were created - that is largely irrelevant.

 

[Taproot]

I've been slowly accumulating too. Roughly what number of shares would be the top 150 or 100. I now have way more than I originally intended.

The last time someone put up a list it was:

300: 525K
250: 625K
200: 800K
150: 1M
100: 1.3M


Might be a little outdated, but should give you a rough guide.

 

[MrNick]

The UK is promoting this... should we? Shouldn't we? Seems tailor-made to park some 'brains' in this new Centre.

UK Multidisciplinary Centre for Neuromorphic Computing

Apply for funding to establish a UK centre for neuromorphic computing. This will be a focal point to engage relevant disciplines, further research and demonstrate the potential of the technology to deliver NetZero solutions.

www.ukri.org