BRN Discussion Ongoing

[Boab]

Anyone have an idea as to who the customer might be that Sean is meeting with in Australia?

I'm buying him lunch and he's going to tell me everything
No Doubt About it.

 

[mcm]

stab in the dark, and most likely wrong!

Real-Time Monitoring and Security Solutions | Spectur

Integrate monitoring, surveillance and warning with Spectur’s AI and IoT-driven solutions. See how our interoperable smart systems can adapt to your needs.

spectur.com.au

Great stab. If they're not talking to Sean ... they bloody well should be!

 

[JDelekto]

Exactly.
And finally answer your legitimate question whether or not Sean Hehir had indeed told select participants of that by-invitation-only shareholder meeting in Sydney in November that Brainchip had succeeded in developing and running a number of LLMs on Akida 2.0 (as claimed by a poster here on TSE), and if true, why the remaining shareholders (which constitute the vast majority) still have not been informed about this amazing breakthrough via official channels four months later.

I also wonder why none of the other attendees of that Sydney gathering have so far shared with us their recollection of what Sean Hehir actually said.

There is something that I would like to point out about LLMs. LLMs can have large memory requirements, depending upon the model quality, the number of parameters, and quantization. For example, a 70-billion parameter CodeLlama Instruct model (useful for writing computer code) with an 8-bit quantization requires about 70+ GB of RAM. You can still get decent results and save space with a lower quantization but it suffers a loss in quality if the quantization is too low.

That same CodeLlama Instruct model with 4-bit quantization is about 40 GB of RAM. I have a desktop with 64 GB and I can partially offload the model to 24 GB of GPU RAM with an Nvidia GTX 3080 card. The more I get off the CPU and onto the video card (which does all the math faster), the faster my responses to chat. Unless one plays a lot of resource-hungry video games or does work with AI, you will not find most consumers that have desktop PCs with that much memory to run those models.

Now, given a base system with a CPU and enough memory to hold a much smaller model (like 7 billion parameters instead of 70) and quantized down to about 4 bits, you're looking at slightly under 4 GB for a model. I think this is something an Edge device can handle and I think that may be useful depending on the use case.

So I think that technically, Akida could run an LLM if it were properly "massaged" (I believe I ran across a GitHub repo where an individual was attempting to run an LLM with a Spiking Neural Network), but running an LLM the size and quality of GPT-4 or Claude-3 would not be practical.

 

[Stable Genius]

ASX fined $1.05m by ASIC, blames 'system configuration error'

ASX Limited has paid a penalty of $1.05 million following an investigation by the Australian Securities and Investments Commission (ASIC) into its compliance with the market integrity rules. It is the first time ASIC has issued an infringement notice to a market operator.

www.capitalbrief.com

 

[overpup]

Fair enough, but I still disagree

The definition of something, with changing relative context, can not remain constant, in my opinion.

Not even sure if that made sense, but it's past even my bedtime...

That old saying attributed to Einstein: "Insanity is doing the same thing over and over and expecting a different result"...
You can tell Albert never worked with computers!

 

[IloveLamp]

 

[Diogenese]

There is something that I would like to point out about LLMs. LLMs can have large memory requirements, depending upon the model quality, the number of parameters, and quantization. For example, a 70-billion parameter CodeLlama Instruct model (useful for writing computer code) with an 8-bit quantization requires about 70+ GB of RAM. You can still get decent results and save space with a lower quantization but it suffers a loss in quality if the quantization is too low.

That same CodeLlama Instruct model with 4-bit quantization is about 40 GB of RAM. I have a desktop with 64 GB and I can partially offload the model to 24 GB of GPU RAM with an Nvidia GTX 3080 card. The more I get off the CPU and onto the video card (which does all the math faster), the faster my responses to chat. Unless one plays a lot of resource-hungry video games or does work with AI, you will not find most consumers that have desktop PCs with that much memory to run those models.

Now, given a base system with a CPU and enough memory to hold a much smaller model (like 7 billion parameters instead of 70) and quantized down to about 4 bits, you're looking at slightly under 4 GB for a model. I think this is something an Edge device can handle and I think that may be useful depending on the use case.

So I think that technically, Akida could run an LLM if it were properly "massaged" (I believe I ran across a GitHub repo where an individual was attempting to run an LLM with a Spiking Neural Network), but running an LLM the size and quality of GPT-4 or Claude-3 would not be practical.

Hi JD,

That's some impressive technowizardry.

As you know, PvdM's "4 Bits are enough" white paper discusses the advantages of 4-bit quantization.

https://brainchip.com/4-bits-are-enough/
...
4-bit network resolution is not unique. Brainchip pioneered this Machine Learning technology as early as 2015 and, through multiple silicon implementations, tested and delivered a commercial offering to the market. Others have recently published papers on its advantages, such as IBM, Stanford University and MIT.

Akida is based on a neuromorphic, event-based, fully digital design with additional convolutional features. The combination of spiking, event-based neurons, and convolutional functions is unique. It offers many advantages, including on-chip learning, small size, sparsity, and power consumption in the microwatt/milliwatt ranges. The underlying technology is not the usual matrix multiplier, but up to a million digital neurons with either 1, 2, or 4-bit synapses. Akida’s extremely efficient event-based neural processor IP is commercially available as a device (AKD1000) and as an IP offering that can be integrated into partner System on Chips (SoC). The hardware can be configured through the MetaTF software, integrated into TensorFlow layers equating up to 5 million filters, thereby simplifying model development, tuning and optimization through popular development platforms like TensorFlow/Keras and Edge Impulse. There are a fast-growing number of models available through the Akida model zoo and the Brainchip ecosystem.

To dive a little bit deeper into the value of 4-bit, in its 2020 NeurIPS paper IBM described the various pieces that are already present and how they come together. They prove the readiness and the benefit through several experiments simulating 4-bit training for a variety of deep-learning models in computer vision, speech, and natural language processing. The results show a minimal loss of accuracy in the models’ overall performance compared with 16-bit deep learning. The results are also more than seven times faster and seven times more energy efficient. And Boris Murmann, a professor at Stanford who was not involved in the research, calls the results exciting. “This advancement opens the door for training in resource-constrained environments,” he says. It would not necessarily make new applications possible, but it would make existing ones faster and less battery-draining“ by a good margin.”

While I have some understanding of the visual aspect, I find the NLP, covering both speech and text, more perplexing mainly because of the need for context or "attention", but, as usual, Prof Wiki has some useful background:

https://en.wikipedia.org/wiki/Natural_language_processing

 

[McHale]

Hi McHale,

I had the same trouble a while ago. Have you tried pressing this icon?

View attachment 58620

Thanks IloveLamp, I will check it out, hmmm

 

[McHale]

I guess he doesn't think BrainChip has reached the "Tipping point" yet then?

And he has an unexplainable "new" ethos?

View attachment 58606

He will do much better for us over there, if he can garner the support and following there, that he once had.

There are many more eyes there, that will notice, if he can manage "Top Rated" posts.

It's a good thing he's changed his approach also, as the smarter trolls there, will tear shreds out of his commentary, if he uses information, the way he did here.

However, personally, I think BrainChip is now beyond the influence of share forums and has reached the tipping point, as far as Global recognition is concerned, anyway..

We are on the World stage now and are rapidly moving, from speculative grade, to investment grade.

A couple of new IP deals, would seal that opinion..

Any potential investors, who find BRN an attractive "punt", influenced by reading "Share Forum" posts, are just fortunate, in my opinion.

All very interesting Mr Dingo, but I thought that FF despised HC, so it's curious he's put up some posts over there (there was another the day before I think), he must've been really annoyed at something.

However FF has a track record of leaving in a huff, if my memory serves me correct it happened twice at HC and now twice here, regardless he is an excellent poster and researcher so it's disappointing he's gone again.

 

[Rskiff]

All very interesting Mr Dingo, but I thought that FF despised HC, so it's curious he's put up some posts over there (there was another the day before I think), he must've been really annoyed at something.

However FF has a track record of leaving in a huff, if my memory serves me correct it happened twice at HC and now twice here, regardless he is an excellent poster and researcher so it's disappointing he's gone again.

Unfortunately with FF posting there, it will be benefitting TMH in terms of traffic to site, thus enriching those fwits. I will stay loyal to Tsex and never go back to that cesspit.

 

[DingoBorat]

All very interesting Mr Dingo, but I thought that FF despised HC, so it's curious he's put up some posts over there (there was another the day before I think), he must've been really annoyed at something.

However FF has a track record of leaving in a huff, if my memory serves me correct it happened twice at HC and now twice here, regardless he is an excellent poster and researcher so it's disappointing he's gone again.

The difference is though, he hasn't left and stopped posting, he's just changed forums.

His integrity was called into question here and instead of being a man and admitting to his mistakes and errors of judgement, he told Frangipani she should "go and get some help"?

And that he felt it no longer necessary to post about BRN, as he had better things to do?

I agree, he is an excellent poster and great researcher, who excels at communication and much more often than not, has a very insightful point of view.
The prodigiousness and quality of his posting, was/is incredible.

I'm not enamoured though, by someone who thinks they are beyond reproach and is not able to handle criticism, or admit fault.

 

[skutza]

The difference is though, he hasn't left and stopped posting, he's just changed forums.

His integrity was called into question here and instead of being a man and admitting to his mistakes and errors of judgement, he told Frangipani she should "go and get some help"?

And that he felt it no longer necessary to post about BRN, as he had better things to do?

I agree, he is an excellent poster and great researcher, who excels at communication and more often than not, has a very insightful point of view.
The prodigiousness and quality of his posting, was incredible.

I'm not enamoured though, by someone who thinks they are beyond reproach and is not able to handle criticism, or admit fault.

Could you imagine living with him?




But yes he does do an excellent job 95% of the time. One of the best I've seen.

 

[IloveLamp]

https://www.reuters.com/world/us/us-officials-face-tough-choices-subsidizing-ai-chip-manufacturing-2024-03-06/

 

[Slade]

Could you imagine living with him?

A Horrible and Dumb thing to say that makes me also feel like I don’t want to spend much time on this forum.

 

[Humble Genius]

All very interesting Mr Dingo, but I thought that FF despised HC, so it's curious he's put up some posts over there (there was another the day before I think), he must've been really annoyed at something.

However FF has a track record of leaving in a huff, if my memory serves me correct it happened twice at HC and now twice here, regardless he is an excellent poster and researcher so it's disappointing he's gone again.

Perhaps he posted it there accidently?

A bit like when you are going to send a text to your girlfriend, and accidently send it to your wife

 

[Makeme 2020]

A Horrible and Dumb thing to say that makes me also feel like I don’t want to spend much time on this forum.

See ya

 

[MDhere]

Ok here i go again @skutza this is an anonymous forum about a stock that is absoulely incredible. its not a forum to judge if one can live with another. Also @DingoBorat , there are two sides to the frangipani story and FF , both had merits but both saw different views and i agree that go and get help wasnt exactly positive it was said in regards to Frangis bible of a story which quite frankly wasnt need on this forum.

Now i admit to watching MAFS but i certainly dont expect my fellow brners to be wondering what underwear i wear and and wether im a suitable companion depending on my texts on here lol

i for one am happy with some long term holders posting on HC as it keeps the stupid site in check a bit (although i have chosen not to but i do frequent it and put like and so forth up) Though at times I want to throttle that Dean guy lol

And lastly @Slade , Mate (sounds so aussie when i write that) Mate don't you dare depart from this forum, or all my beer shouting is off! lol

On a side not i can't get long skip connections out of my head i think also known as long residual connections as Sean has mentioned this a number of times stating "our customers" asked for this and we have delivered what they are asking for.

To me that is like a medical purpose this long skip connection. The other thing was somatosensory which I believe will bring up assistance to articial limbs momement. and the US Department of Homeland Security have been playing iwth Adika in the past when i googled about the skp/residual connections.

O and FF said something about Ericsson, I don't see it on our iceberg nor on our resident flowcharting pencil pusher A4 paper detective (cant remember is that you @Esq.111 ) or it might be someone else but i loved it and prob needs an update?

Cheers fellow brners Thats it for me for now, peace to all and I will no doubt re-appear tomorrow as we are all strapped in

 

[CHIPS]

Perhaps he posted it there accidently?

A bit like when you are going to send a text to your girlfriend, and accidently send it to your wife

I know a worse situation which happened to a girlfriend of mine. Her husband sent a WhatApp to his girlfriend talking about the great night they had, well at least he tried to, but by accident, he sent it to his family group, i.e. his wife and three teenage kids. Needless to say, they are getting a divorce now.
First, she was sad, upset, angry ... but a few days later she asked me: What is better than marrying a rich man? Getting divorced from a rich man!!!

 

[Esq.111]

Afternoon MDhere ,

Have not added ERICSSON yet ... though thay have / are playing with our tech , as per the documents reveiled.

Need a little more time and proof...confirmation from their or our chief Indian.

THE BRAINCHIP SCROLL.

Note : there may well be others which I have missed ... so this should by no means be considered exhaustive.

Regards,
Esq.

 

[Tothemoon24]

Interesting development.​

If this is to be accurate, & our lead in the neuromorphic SNN space is correct ….?

Then we would be part of this ….?

If this is accurate & we are not part of it ,…

​

KAIST researchers develop world's first 'neuromorphic' AI chip

A research team at KAIST has developed the world’s first AI semiconductor capable of processing a large language model (LLM) with ultra-low power consumption, the Science Ministry said Wednesday.

koreajoongangdaily.joins.com

KAIST researchers develop world's first 'neuromorphic' AI chip​

Neuromorphic computing technology aims to develop integrated circuits mimicking the human nervous system so that chips could be able to perform more sophisticated tasks. [SHUTTERSTOCK]
A research team at KAIST has developed the world’s first AI semiconductor capable of processing a large language model (LLM) with ultra-low power consumption using neuromorphic computing technology.

The technology aims to develop integrated circuits mimicking the human nervous system so that chips could be able to perform more sophisticated tasks that require adaption and reasoning with far less energy consumption.


The Science Ministry said Wednesday that the team, led by Prof. Yoo Hoi-jun at the KAIST PIM Semiconductor Research Center, developed a “Complementary-Transformer” AI chip, which processes GPT-2 with an ultra-low power consumption of 400 milliwatts and a high speed of 0.4 seconds, according to the Ministry of Science and ICT.

A rendered image of comparing performance of different types of processors [YONHAP]

The 4.5-millimeter-square chip, developed using Korean tech giant Samsung Electronics' 28 nanometer process, has 625 times less power consumption compared with global AI chip giant Nvidia’s A-100 GPU, which requires 250 watts of power to process LLMs, the ministry explained.

The chip is also 41 times smaller in area than the Nvidia model, enabling it to be used on devices like mobile phones, therefore better protecting user privacy.

The KAIST team has succeeded in demonstrating various language processing tasks with its LLM accelerator on Samsung’s latest smartphone model, the Galaxy S24, which is the world’s first smartphone model with on-device AI, featuring real-time translation for phone calls and improved camera performance, Kim Sang-yeob, a researcher on the team, told reporters in a press briefing.

The ministry said the utilization of neuromorphic computing technology, which functions like a human brain, specifically spiking neural networks (SNNs), is essential to the achievement.

Previously, the technology was less accurate than deep neural networks (DNNs) and mainly capable of simple image classifications, but the research team succeeded in improving the accuracy of the technology to match that of DNNs to apply it to LLMs.

The team said its new AI chip optimizes computational energy consumption while maintaining accuracy by using unique neural network architecture that fuses DNNs and SNNs and effectively compresses the large parameters of LLMs.

“Neuromorphic computing is a technology global tech giants, like IBM and Intel, failed to realize. We believe we are the first to run an LLM with a ultra-low power neuromorphic accelerator,” Yoo said.

BY PARK EUN-JEE, YONHAP [park.eunjee@joongang.co.kr]