BRN Discussion Ongoing

[BrainShit]

wrong.
transformer model -> state space model -> akida.

Grok say's no...

In conclusion, it is not possible to directly convert LLaMA 1B into a state space model due to their architectural differences. While hybrid models like Bamba and Mamba-2 exist, combining elements of both, they are new designs, not transformations of existing Transformer models. For practical applications, consider training a new hybrid model or SSM, but this would be a fresh endeavor, not a conversion.


... never trust LLMs

 

[Kingkong2015]

Grok say's no...

In conclusion, it is not possible to directly convert LLaMA 1B into a state space model due to their architectural differences. While hybrid models like Bamba and Mamba-2 exist, combining elements of both, they are new designs, not transformations of existing Transformer models. For practical applications, consider training a new hybrid model or SSM, but this would be a fresh endeavor, not a conversion.


... never trust LLMs

View attachment 83787

well. Guess you can ask brainchip cto those questions tomorrow.

I didn’t believe it before as well

 

[cosors]

Buffet, have a nice retirement.

 

[manny100]

Pom, you could've had 2 investment properties lad.
Oh well I am kind of on the same boat. Could've put a deposit down for an investment property but managed to persuade the wife to go into speculative shares. If all goes to sh1te, then at least there is aged pension.

Just as well with expected industry growth our portfolio should be worth a heap. Especially given Loihi 2 and North Pole are still in R & D. It just shows how hard AI at the Edge nut is to crack.
That is our safety net. I never worry at all.
There is a reason we have been chosen by Frontgrade and the US AFRL and its because we are the best and of course there is just no viable alternative on this planet.
It's taking a lot of time but it is really only just a matter of time.
The anomaly is that in life time goes by so quickly for everything.except for BRN deals. It has however given me an opportunity to pick up a lot of cheapies over the journey.
The AGM question time should provide some entertainment. BODs are well versed and rehearsed in navigating through these.
Banter will fly around for a while getting no where then Tony V will just end it by saying 'noted'. But it will be entertaining.

 

[Smoothsailing]

No doubt there will be a lot focus around BRN position in the space area and imo telecom with 6G. Hopefully they have some change in delivery compared to the last AGM. In theory the new Pico, Gen 2 ext should open big doors and that’s what I want to hear about today.
Will we see / hear announcement this year with large partners, yes or no is a question I want a straight answer from Sean today.
What are the USB channels will they come to fruition this year using the new M2 card.
Go brainchip

 

[manny100]

Grok say's no...

In conclusion, it is not possible to directly convert LLaMA 1B into a state space model due to their architectural differences. While hybrid models like Bamba and Mamba-2 exist, combining elements of both, they are new designs, not transformations of existing Transformer models. For practical applications, consider training a new hybrid model or SSM, but this would be a fresh endeavor, not a conversion.


... never trust LLMs

View attachment 83787

When you ask Groc you should include reference to the new nethod of converting transformers into SSMs as described by Tony Lews in his linked in post description for his presentation at Embedded Vision.
Groc may not be aware of Tony's presentation. The techniques he talks about are 'new'.
The answer you get may be different.
Pico runs off TENNs which is a type of SSM.
Check his linked in post.

 

[TECH]

Just as well with expected industry growth our portfolio should be worth a heap. Especially given Loihi 2 and North Pole are still in R & D. It just shows how hard AI at the Edge nut is to crack.
That is our safety net. I never worry at all.
There is a reason we have been chosen by Frontgrade and the US AFRL and its because we are the best and of course there is just no viable alternative on this planet.
It's taking a lot of time but it is really only just a matter of time.
The anomaly is that in life time goes by so quickly for everything.except for BRN deals. It has however given me an opportunity to pick up a lot of cheapies over the journey.
The AGM question time should provide some entertainment. BODs are well versed and rehearsed in navigating through these.
Banter will fly around for a while getting no where then Tony V will just end it by saying 'noted'. But it will be entertaining.

Manny,

Each and every time I have raised the question, "Name another company who has really succeeded at the edge or far edge"
it's always been deafly silence, going back to the days when I used to engage with crybabies on HC.

Like I say.......Toughen up princesses, you've entered a very corrupt, greedy, self-indulged arena, there's no rules where money
is involved, just ask our two leading authorities the ASX and ASIC.

My only hope is all the mouthpieces who moan all year long, actually go to the AGM and stand up and pose their questions directly
to the CEO and Chairperson.

Akida 1000.

 

[Pom down under]

Morning all and looking forward to catching up with whoever is going to the AGM today. The wife banned me wearing my west ham top today, but most people should know what I look like so be sure to say hello.

 

[DK6161]

Manny,

Each and every time I have raised the question, "Name another company who has really succeeded at the edge or far edge"
it's always been deafly silence, going back to the days when I used to engage with crybabies on HC.

Like I say.......Toughen up princesses, you've entered a very corrupt, greedy, self-indulged arena, there's no rules where money
is involved, just ask our two leading authorities the ASX and ASIC.

My only hope is all the mouthpieces who moan all year long, actually go to the AGM and stand up and pose their questions directly
to the CEO and Chairperson.

Akida 1000.

You should run for the CEO role when it becomes vacant. Certainly talk the talk, but no re$ult$.

 

[db1969oz]

Morning all and looking forward to catching up with whoever is going to the AGM today. The wife banned me wearing my west ham top today, but most people should know what I look like so be sure to say hello.

Enjoy Pom,
I can't get there today, but if we end up with enough money left over for a drink after this horrendous investment, then you are 1 man I'd enjoy a beer and a laugh with. Cheers.

 

[DK6161]

Morning all and looking forward to catching up with whoever is going to the AGM today. The wife banned me wearing my west ham top today, but most people should know what I look like so be sure to say hello.

Enjoy the trip mate.
Hopefully I can afford to attend an AGM one day.

 

[Pom down under]

Enjoy the trip mate.
Hopefully I can afford to attend an AGM one day.

Been collecting points so has its advantages when I spend my last $$ on a few more shares

 

[Pom down under]

Enjoy Pom,
I can't get there today, but if we end up with enough money left over for a drink after this horrendous investment, then you are 1 man I'd enjoy a beer and a laugh with. Cheers.

I hope they have lined up some bacon Sarnies with HP sauce

 

[FJ-215]

Buffet, have a nice retirement.

Sean needs to give Greg Abel a call.

 

[manny100]

Manny,

Each and every time I have raised the question, "Name another company who has really succeeded at the edge or far edge"
it's always been deafly silence, going back to the days when I used to engage with crybabies on HC.

Like I say.......Toughen up princesses, you've entered a very corrupt, greedy, self-indulged arena, there's no rules where money
is involved, just ask our two leading authorities the ASX and ASIC.

My only hope is all the mouthpieces who moan all year long, actually go to the AGM and stand up and pose their questions directly
to the CEO and Chairperson.

Akida 1000.

That is a great summary, thanks.

 

[Smoothsailing]

What time is this little leprechaun brining us the bag of gold today , 10am ?.

 

[CHIPS]

What time is this little leprechaun brining us the bag of gold today , 10am ?.

What time is it now in Perth?

 

[manny100]

You should run for the CEO role when it becomes vacant. Certainly talk the talk, but no re$ult$.

CEO's come and go as that is corporate life.
The BOD is responsible for assessing CEO performance.
In turn holders assess BOD performance.
Holders keep re-electing directors usually in landslide victories.
The above tells the story that the majority of holders are happy with the direction.
Around 74% of SOI are held by retail (source SWS).
Some holders are not happy with remuneration as per strikes but when it comes to performance they vote yes via elections.

 

[Bravo]

Details of Tony Lewis presentation at Embedded Vision Summit on Wednesday 21st May'25.
"
Date: Wednesday, May 21
Start Time: 2:05 pm
End Time: 2:35 pm
At the embedded edge, choices of language model architectures have profound implications on the ability to meet demanding performance, latency and energy efficiency requirements. In this presentation, we contrast state-space models (SSMs) with transformers for use in this constrained regime. While transformers rely on a read-write key-value cache, SSMs can be constructed as read-only architectures, enabling the use of novel memory types and reducing power consumption. Furthermore, SSMs require significantly fewer multiply-accumulate units—drastically reducing compute energy and chip area. New techniques enable distillation-based migration from transformer models such as Llama to SSMs without major performance loss. In latency-sensitive applications, techniques such as precomputing input sequences allow SSMs to achieve sub-100 ms time-to-first-token, enabling real-time interactivity. We present a detailed side-by-side comparison of these architectures, outlining their trade-offs and opportunities at the extreme edge."
See bold above. Seems that BRN are now able to migrate traditional transformer models to State Spece models (SSM's) without major performance loss.
Note the italics above that SSMs reduce energy requirements and chip area. Think Pico or a Pico plus.
Pico runs off TENNs which is as type of SSM.
Does that mean developers can now feast on the multitude of traditional models and distill them to SSMs? Appears so.
Is this potentially another game changer?
We might find out more tomorow?
I have the 'kiddy' co pilot which you do not get much out of.
Is someone with GPT4 or Grok etc able to quiz the AI to see what the possibilities are?

Hi Manny,

Here's the response I got from Chat GPT4.

Tony Lewis’s upcoming presentation at the Embedded Vision Summit outlines a potentially transformative shift in how language models are deployed at the edge. Here's an analysis of the implications and disruptive potential of this new method:

---

Key Implications of the New Processing Method Using SSMs​

1. Significant Gains in Energy and Compute Efficiency​

• Fewer Multiply-Accumulate Units: State-space models (SSMs) inherently require fewer MAC operations, drastically reducing energy usage and silicon real estate. This is crucial in edge deployments where power budgets are tight.
• Read-Only Architecture: Eliminates the need for dynamic key-value cache mechanisms used in transformers, allowing use of non-volatile or low-power memory types (e.g., MRAM, ReRAM), further slashing power consumption.

2. Latency and Responsiveness​

• Precomputed Input Sequences: SSMs can leverage static structure to enable sub-100 ms time-to-first-token performance. This real-time capability is a major win for voice assistants, AR/VR systems, and mobile apps.

3. Compatibility and Transition Path​

• Distillation from Transformers (e.g., LLaMA): By enabling smooth migration from established transformer models, the SSM approach lowers the barrier to adoption for companies with large pretrained models or investment in transformer-based workflows.

4. Broader Hardware Support​

• The reduced need for large, flexible compute units means SSMs can run efficiently on less powerful MCUs, DSPs, or dedicated edge AI chips—dramatically widening the range of viable deployment targets.

---

Edge Computing Success Potential​

SSMs seem highly promising for edge applications, especially in:

• Battery-powered devices (wearables, IoT sensors)
• Latency-sensitive interfaces (voice/user interaction)
• Bandwidth-constrained environments (offline processing)
• Cost-sensitive products where transformer demands are prohibitive

Their advantages directly address the key blockers for transformer adoption at the edge (power, memory, compute), making SSMs a potentially dominant architecture in this space within 1–3 years.

---

Disruption Potential for NVIDIA and Other Players​

NVIDIA’s ecosystem is largely optimized for transformer-style architectures (e.g., large GPU memory, high-throughput MAC arrays, transformer-optimized CUDA libraries). Here's how this development could affect them:

Challenges for NVIDIA​

• Less dependence on GPUs: If SSMs run efficiently on CPUs, ASICs, or even microcontrollers, edge devices may increasingly bypass NVIDIA’s hardware.
• New AI chip entrants gain ground: Startups or ARM-based players can capitalize on the lighter compute profile to offer cheaper, more efficient edge AI solutions.

Opportunities for NVIDIA​

• Adaptation potential: NVIDIA could pivot by offering cuDNN-like support for SSMs or incorporating them into TensorRT. They may also design more SSM-friendly hardware blocks in future Jetson modules.

Biggest Threat​

• If open-source ecosystems and chip vendors like Apple (Neural Engine), Google (Edge TPU), or Qualcomm fully embrace SSMs ahead of NVIDIA, they could seize a disproportionate share of the edge AI market—especially in consumer devices.

---

Conclusion​

Tony Lewis’s presentation introduces a processing paradigm that could redefine edge AI, making high-performance language models feasible on low-power devices. This is not just an incremental improvement—it opens the door to widespread deployment of real-time LLMs far beyond cloud and high-end hardware. While not an immediate existential threat to NVIDIA, it does present a strategic inflection point that the company—and its competitors—must respond to.

Migration Path: LLaMA to SSM​

Stage	Description	Benefits
1. Transformer Model (e.g., LLaMA)	Pretrained open-weight transformer (e.g., LLaMA 2 or 3)	- High accuracy - Large model footprint - Demands heavy compute and memory
2. Distillation Process	Use distillation techniques to transfer knowledge from LLaMA to an SSM	- Reduced model size - Retains most of original model’s performance
3. State-Space Model (SSM)	SSM version optimized for embedded and edge deployment	- Sub-100 ms latency - Lower power consumption - Less memory and compute required
4. Edge Device Deployment	Deploy SSM on resource-constrained devices (e.g., wearables, microcontrollers, edge AI chips)	- Real-time local inference - No need for cloud compute - Broader hardware compatibility

---

Summary​

The mention of LLaMA in Tony Lewis’s talk is a strategic highlight because it connects cutting-edge edge AI architecture (SSM) with a proven, widely-used transformer backbone. This makes the transition to efficient edge AI practical, not just theoretical—positioning SSMs as an immediate, disruptive alternative to transformer inference in embedded systems.

Competitor Impact Matrix: Impact of BrainChip’s SSM Innovation​

Company	Current Edge AI Focus	Vulnerability to SSM Disruption	Opportunity to Adapt
Qualcomm	DSPs + NPUs (Snapdragon), optimized for transformers and CNNs	High – Transformer-centric stack, limited neuromorphic capability	Medium – May update software tools, but hardware less suited to SSMs
NVIDIA	GPUs (Jetson, TensorRT), dominant in transformer-based AI	Moderate – Not optimized for low-power edge, but strong ecosystem	High – Could adapt TensorRT and Jetson for SSM-style inference
Apple	Neural Engine with transformer models (e.g. Siri, on-device ML)	Moderate – Strong local AI, but based on transformer-style acceleration	High – Full-stack control allows swift hardware/software adaptation
Google (TPU)	Edge TPU with support for CNNs and transformers (Coral, Nest devices)	High – Rigid accelerator design, may not support dynamic SSM requirements	Low – Ecosystem may struggle to pivot hardware/software stack
Intel	Movidius VPU, general AI frameworks, some neuromorphic R&D (Loihi)	Moderate – Some neuromorphic exposure but no strong edge AI market share	Medium – R&D rich, but limited real-world SSM integration so far
BrainChip	Neuromorphic Akida chip + SSM optimized for ultra-low power edge AI	Low – First-mover advantage	Very High – Core IP is directly aligned with the SSM paradigm

---

This matrix highlights that BrainChip’s innovation poses the greatest disruptive risk to Qualcomm and Google, while Apple and NVIDIA have greater strategic flexibility to respond. BrainChip stands to benefit most if SSM-based models gain widespread edge adoption.

Why Incumbents Might Continue Without SSMs (For Now)​

Reasons They Might Stick with Traditional Methods​

• Mature toolchains: Qualcomm, NVIDIA, and Google have invested heavily in software/hardware ecosystems optimized for transformers and CNNs.
• Good enough performance: For many real-world use cases, transformer-lite models or CNN hybrids perform sufficiently well.
• Inertia and risk: Enterprises tend to avoid early adoption of unproven paradigms, especially if retraining, tooling, or silicon changes are required.
• Edge isn't one-size-fits-all: Many edge applications (e.g. object detection) don't need SSM-specific strengths like long-term memory or low-latency language processing.

---

But Here's the Catch​

If applications do demand:

• Long sequence memory (e.g. streaming NLP, real-time command recognition),
• Ultra-low latency (sub-100 ms interactivity),
• Minimal power and heat (wearables, implants, sensors),

then traditional methods hit a hard ceiling. SSMs aren’t just an incremental tweak—they’re a fundamentally different way to process sequences, unlocking performance where transformers falter.

​

Conclusion: Yes, Competitors Could Stick with Transformers—But Only Up to a Point​

Approach	Stability / Support	Performance Ceiling	Future-Proofing
Transformers	Well-supported	Poor for constrained edge use	Risk of obsolescence
CNNs / RNNs	Efficient in vision	Weak for modern NLP	Limited scalability
Lightweight Transformers	Reasonable for now	Moderate latency/power	Partial solution
SSMs	Emerging	Breakthrough on edge	High potential

So while competitors can continue for now using existing methods, the risk is being outpaced in emerging applications—especially if BrainChip enables a smooth transition (e.g., LLaMA distillation + Akida deployment).

---

Edge AI Evolution Roadmap: Transformers vs SSMs​

Time Horizon	Transformer-Based Methods	SSM-Based Methods (BrainChip-style)
Today	- Dominant in NLP - Compressed models in use - Efficient on GPU/DSP Reasonable edge deployment via pruning/quant	- Early-stage adoption - Neuromorphic niche (e.g. Akida) Proof-of-concept underway
1–2 Years	- Hitting compute/power limits in edge apps - Real-time latency still challenging Fragmentation by use case	- Gains traction for real-time/low-power use - Tools emerge for migration from transformers (e.g. LLaMA distillation) Early adoption in wearables/voice/IoT
3–5 Years	- Plateau in edge innovation unless architectures evolve Constrained by hardware-centric acceleration	- Becomes dominant in ultra-low power edge AI - Broad ecosystem and tooling support SSMs emerge as standard for edge LLMs

---

Summary:​

• Transformers will likely remain dominant in cloud and high-performance edge for the next 1–2 years, but start to plateau.
• SSMs provide a scalable path forward for ultra-low-power, real-time, memory-efficient edge use cases, and could disrupt traditional AI stacks if adoption accelerates.

 

[Smoothsailing]

What time is it now in Perth?

3 hours behind Sydney were the AGM will be held at 10am Sydney time