BRN Discussion Ongoing

[Fullmoonfever]

New GitHub update 20hrs ago on Akida/CNN2SNN and including TENNs release, modules and models etc by the looks.

@Diogenese will probs know more of anything unusual or new tucked in there.

Releases · Brainchip-Inc/akida_examples

Brainchip Akida Neuromorphic System-on-Chip examples and documentation. - Brainchip-Inc/akida_examples

github.com

20 hours ago

ktsiknos-brainchip
2.13.0-doc-1
d8435c2
Compare
Upgrade to Quantizeml 0.16.0, Akida/CNN2SNN 2.13.0 and Akida models 1.7.0
Latest

Update QuantizeML to version 0.16.0​

New features​

• Added a bunch of sanitizing steps targetting native hardware compatibility:
  • Handle first convolution that cannot be a split layer
  • Added support for "Add > ReLU > GAP" pattern
  • Added identity layers when no merge layers are present after skip connections
  • BatchNormalisation layers are now properly folded in ConvTranspose nodes
  • Added identity layers to enforce layers to have 2 outbounds only
  • Handled Concatenate node with a duplicated input
• Added support for TENNs ONNX models, which include sanitizing, converting to inference mode and quantizing
• Set explicit ONNXScript requirement to 0.2.5 to prevent later versions that use numpy 2.x

Bug fixes​

• Fixed an issue where calling sanitize twice (or sanitize then quantize) would lead to invalid ONNX graphs
• Fixed an issue where sanitizing could lead to invalid shapes for ONNX Matmul/GEMM quantization

Update Akida and CNN2SNN to version 2.13.0​

Aligned with FPGA-1679(2-nodes)/1678(6-nodes)​

New features​

• [cnn2snn] Updated requirement to QuantizeML 0.16.0
• [cnn2snn] Added support for ONNX QuantizedBufferTempConv and QuantizedDepthwiseBufferTempConv conversion to Akida
• [akida] Full support for TNP-B in hardware, including partial reconfiguration with a constraint that TNP-B cannot be the first layer of a pass
• [akida] Full support of Concatenate layers in hardware, feature set aligned on Add layers
• [akida] Prevented the mapping of models with both TNP-B and skip connections
• [akida] Renamed akida.NP.Mapping to akida.NP.Component
• [akida] Improved model summary for skip connections and TNP-B layers. The summary now shows the number of required SkipDMA channels and the number of components by type.
• [akida] Updated mapping details retrieval: model summary now contains information on external memory used. For that purpose, some C++/Python binding was updated and cleaned. The NP objects in the API have external members for memory.
• [akida] Renamed existing virtual devices and added SixNodesIPv2 and TwoNodesIPv2 devices
• [akida] Introduced create_device helper to build custom virtual devices
• [akida] Mesh now needs an IP version to be built
• [akida] Simplified model statistics API and enriched with inference and program clocks when available
• [akida] Dropped the deprecated evaluate_sparsity tool

Update Akida models to 1.7.0​

• Updated QuantizeML dependency to 0.16.0 and CNN2SNN to 2.13.0
• Sparsity tool name updated. Now returns python objects instead of simply displaying data and support models with skip connections
• Introduced tenn_spatiotemporal submodule that contains model definition and training pipelines for DVS128, EyeTracking and Jester TENNs models
• Added creation and training/evaluation CLI entry points for TENNs

Introducing TENNs modules 0.1.0​

• First release of the package that aims at providing modules for Branchip TENNs
• Contains blocks of layers for model definition: SpatialBlock, TemporalBlock, SpatioTemporalBlock that come with compatibility checks and custom padding for Akida
• The TemporalBlock can optionally be defined as a PleiadesLayer following https://arxiv.org/abs/2405.12179
• An export_to_onnx helper is provided for convenience

Documentation update​

• Added documentation for TENNs APIs, including tenns_modules package
• Introduced two spatiotemporal TENNs tutorials
• Updated model zoo page with mAP50, removed 'example' column and added TENNs
• Added automatic checks for broken external links and fixed a few
• Cosmetic changes: updated main logo and copyright to 2025

 

[Diogenese]

Hmmmm...

Intel’s is overhauling its engineering leadership team in an effort to help with its AI comeback. New hires include executives such as Jean-Didier Allegrucci (AI SoC) and Shailendra Desai (AI architecture) signals a pivot toward AI-first development, with neuromorphic computing potentially in the spotlight.



EXTRACT 1
View attachment 87300


EXTRACT 2
View attachment 87301

Intel Starts Re-Engineering Its Executive Ranks

It has been two and a half months since new chief executive officer Lip-Bu Tan gave the keynote at Intel’s Vision 2025 event, and the company has been

www.nextplatform.com

Intel Revamps Leadership with Tech Giants to Boost AI Chip Strategy

Intel Revamps Leadership with Tech Giants to Boost AI Chip Strategy

www.ainvest.com

Intel Appoints Sales and Engineering Leaders - Intel Newsroom

New appointments align with focus on delivering for customers and becoming an engineering-focused company.

newsroom.intel.com

EXTRACT

View attachment 87302
In the news today,

Apparently Altman does not believe in the once bitten, twice shy adage.

When he first pledged $50M to Rain AI, their proposed product, confusing complexity with chaos, was a spaghetti bowl of "self-organizing" nanowires which were supposed to autonomously establish analog neuronal connexions.

A couple of years down the track, they abandoned this and switched to digital neurons. Now they appear to be workiing on hybrid digital/analog with ADC/DAC

Still, as they say, another $150M is neither here nor there.

https://finance.yahoo.com/news/sam-...TPcs9cWc0tQ7adDpJ0Knav5r36EtB13T-ts0zmG9uIl0Z

20250524​

Sam Altman's $150M AI Chip Bet Crashes: Rain AI Faces Sale As OpenAI, Nvidia, And Microsoft Circle The Wreckage​

 

[Diogenese]

An EXTRACT from this evenings press release.


View attachment 87322

Launch of BrainChip Developer Hub Accelerates Event-Based AI Innovation on Akida™ Platform with Release of MetaTF 2.13

BrainChip announces new Developer Hub and MetaTF toolkit, enabling seamless development and deployment of machine learning models on its Akida™ platform.

www.einpresswire.com

Hopefully...

View attachment 87323
View attachment 87324

It will be interesting to see some power usage figures for the glasses.

 

[Frangipani]

The 40th Space Symposium in Colorado Springs is in full swing.

Josef Aschbacher, the European Space Agency’s Director General, reminisces about a warm summer night in his childhood that kindled the flame of curiosity in him and laid the foundation for his life-long fascination with space.

In his post, he also shares a link to ESA’s Strategy 2040, of which you’ll find a summary below:

We all have our own memories deep in our subconscious that can catch us off-guard when suddenly evoked by a distinct smell, a familiar voice, the mention of an old friend we haven’t thought of in… | Josef Aschbacher | 12 comments

We all have our own memories deep in our subconscious that can catch us off-guard when suddenly evoked by a distinct smell, a familiar voice, the mention of an old friend we haven’t thought of in years. Where we can remember exactly what we were doing, wearing, thinking, feeling. Viscerally, we...

www.linkedin.com

View attachment 81805

https://esamultimedia.esa.int/docs/corporate/ESA_Strategy_2040_InBrief.pdf

(For the Strategy 2040 In Focus and Strategy 2040 In Depth versions, click here)

View attachment 81800 View attachment 81801 View attachment 81802 View attachment 81803 View attachment 81804

On Day 2 of the Paris Air Show 2025 at Le Bourget, ESA released Technology 2040, their vision for the next 15 years and beyond, which “fully supports and enhances” ESA’s Strategy 2040, published earlier this year .

Laurent Hili is named as one of the contact persons for the subtopic “AI-DRIVEN INNOVATIONS IN SPACE TECHNOLOGIES”.

ESA Technology 2040 is not just a roadmap – it is a call to action. The document defines an integrated tech stack that keeps supply‑chains sovereign, scales through serialisation and modularity, and… | Josef Aschbacher | 23 comments

ESA Technology 2040 is not just a roadmap – it is a call to action. The document defines an integrated tech stack that keeps supply‑chains sovereign, scales through serialisation and modularity, and locks in European leadership from LEO to deep space. By 2040, we envision a resilient European...

www.linkedin.com

https://esamultimedia.esa.int/docs/technology/Technology_2040.pdf

Some excerpts:

 

[Pom down under]

Hire The Best Crypto Recovery Specialists: HIRE CERTIFIED RECOVERY SERVICES

I met someone on a kink connect app, and our conversations quickly escalated. The chemistry was palpable, and what started as casual chats soon turned into deep discussions about our desires and fantasies. The individual seemed knowledgeable and charismatic, which made me feel at ease. As we exchanged messages, I found myself drawn into their world, captivated by their charm and the promise of an exciting connection. Our relationship developed, the conversation took a surprising turn. The person introduced the idea of investing in cryptocurrency, claiming it was a fantastic opportunity to grow wealth quickly. At first, I had heard about the volatility of crypto markets and the risks involved. But the allure of potential profits, combined with the trust I had built in our interactions, made me reconsider. The scammer provided what appeared to be legitimate instructions for transferring my crypto assets to a banking app based in Singapore. They assured me that this was a safe and secure method to invest, emphasizing the potential returns I could reap. Trusting their guidance, I decided to take the plunge. I sent Ethereum (ETH), Bitcoin (BTC), and USDT in three separate transactions, each time feeling a mix of excitement and apprehension. I believed I was making a wise financial decision, convinced that I was on the brink of a lucrative investment. It wasn’t until I attempted to withdraw my supposed profits that the reality of the situation hit me. The app became unresponsive, and my attempts to contact the scammer went unanswered. Panic set in as I realized I had fallen victim to a sophisticated scam. I had lost thousands of dollars, and the weight of that loss was crushing. Desperate to recover my funds, I turned to CERTIFIED RECOVERY SERVICES, a company specializing in cryptocurrency recovery. Their expertise in tracing lost assets gave me hope. The team at CERTIFIED RECOVERY SERVICES conducted a thorough blockchain forensic analysis and was able to trace the scammer’s wallet. This intricate process required specialized knowledge, and I was grateful to have their support in navigating this challenging situation. After several days of investigation and communication with various financial institutions, CERTIFIED RECOVERY SERVICES successfully recovered all my funds, totaling approximately $287,000. The relief I felt was overwhelming; I had not only regained my financial stability but also learned a valuable lesson about the importance of vigilance in online interactions. This has made me more cautious in my online dealings. While the world of online dating can be exciting, it is crucial to remain aware of the potential risks, especially when financial transactions are involved. Thanks to CERTIFIED RECOVERY SERVICES, I was able to reclaim my lost assets.
Here's Their Info Below:
WhatsApp: (+1(740)258‑1417 )
Telegram: https: //t.me/certifiedrecoveryservices
mail: (certifiedrecoveryservices @zohomail .com, certified @financier .com)
Website info;( https: //certifiedrecoveryservices .com)

@zeeb0t

 

[Rach2512]

@zeeb0t

That went straight on ignore.

 

[7für7]

Smells like a big announcement is right on the corner!!
It smells nice… like a fresh warm wind after a long and hard winter…
It’s still March, but there’s something in the air… something whispering softly:
“Hey… don’t worry… the climate change will bring me back earlier than usual…
Just relax… don’t worry… you don’t have to wait much longer…”
…with a calm female voice… clean, mysterious… a bit seductive, too.
Not too pushy.. just enough to make you feel like she knows something.
And maybe… just maybe… she’s already here. But she wants us to feel it before we see it.

Ahhh… the announcement… I can feel it… like a new season coming, even before the calendar allows it.
Yes… it’s coming. Right? …Right??
Or maybe…
…it’s just my radiator kicking in again.
Hard to tell these days.

 

[Gazzafish]

https://www.dqindia.com/esdm/brainchip-collaborates-with-chelpis-mirle-on-security-solution-9370377

Extract-“

BrainChip collaborates with Chelpis-Mirle on security solution​

M.2 card is based on a design from BrainChip along with an agreement to purchase a significant number of AKD1000 chips for qualification and deployment.​

17 Jun 2025 13:10 IST

DQI Bureau
Follow Us

Listen to this article
BrainChip Holdings Ltd, the world’s first commercial producer of ultra-low power, fully digital, event-based, neuromorphic AI, announced that Chelpis Quantum Corp. has selected its Akida AKD1000 chips to serve as the processor for built-in post-quantum cryptographic security.

Chelpis, a chip company leading the Quantum Safe Migration ecosystem in Taiwan, is developing an M.2 card using the AKD1000 that can be inserted into targeted products to support their cryptographic security solutions. The M.2 card is based on a design from BrainChip along with an agreement to purchase a significant number of AKD1000 chips for qualification and deployment. Upon completion of this phase, Chelpis is planning to increase its commitment with additional orders for the AKD1000.
This agreement is the first step in a collaboration that is exploring the development of an AI-PQC robotic chip designed to fulfill both next-generation security and AI computing requirements. This project is a joint development effort with Chelpis partner company Mirle (2464.TW) and has been formally submitted for consideration under Taiwan’s chip innovation program. The funding aims to promote a new system-on-chip (SoC) integrating RISC-V, PQC, and NPU technologies.
This SoC will specifically support manufacturing markets that emphasize a Made-in-USA strategy. Mirle plans to build autonomous quadruped robotics that mimic the movement of four-legged animals for industrial/factory environments. To enable this vision, Chelpis is exploring BrainChip’s advanced Akida™ IP to incorporate advanced visual GenAI capabilities in the proposed SoC design.

“The ability to add Edge AI security capabilities to our industrial robotics project that provides the low power data processing required is paramount to successfully achieving market validation in the robotics sector,” said Ming Chih, CEO of Chelpis. “We believe that BrainChip’s Akida is just the solution that we further need to bring our SoC to fruition. Their event-based processing and advanced models serve as a strong foundation for developing a platform for manufacturing customers looking to leverage advanced robotics in their facilities.”
“Akida’s ability to efficiently provide cyber-security acceleration with energy efficiency can help secure autonomous robotic devices,” said Sean Hehir, CEO of BrainChip.
“Akida’s innovative approach to supporting LLMs and GenAI algorithms could serve as a key contributor to Chelpis as they pursue government funding to develop their SoC and advance their industrial robotic initiatives.”

 

[7für7]

https://www.dqindia.com/esdm/brainchip-collaborates-with-chelpis-mirle-on-security-solution-9370377

Extract-“

BrainChip collaborates with Chelpis-Mirle on security solution​

M.2 card is based on a design from BrainChip along with an agreement to purchase a significant number of AKD1000 chips for qualification and deployment.​

17 Jun 2025 13:10 IST

DQI Bureau
Follow Us

Listen to this article
BrainChip Holdings Ltd, the world’s first commercial producer of ultra-low power, fully digital, event-based, neuromorphic AI, announced that Chelpis Quantum Corp. has selected its Akida AKD1000 chips to serve as the processor for built-in post-quantum cryptographic security.

Chelpis, a chip company leading the Quantum Safe Migration ecosystem in Taiwan, is developing an M.2 card using the AKD1000 that can be inserted into targeted products to support their cryptographic security solutions. The M.2 card is based on a design from BrainChip along with an agreement to purchase a significant number of AKD1000 chips for qualification and deployment. Upon completion of this phase, Chelpis is planning to increase its commitment with additional orders for the AKD1000.
This agreement is the first step in a collaboration that is exploring the development of an AI-PQC robotic chip designed to fulfill both next-generation security and AI computing requirements. This project is a joint development effort with Chelpis partner company Mirle (2464.TW) and has been formally submitted for consideration under Taiwan’s chip innovation program. The funding aims to promote a new system-on-chip (SoC) integrating RISC-V, PQC, and NPU technologies.
This SoC will specifically support manufacturing markets that emphasize a Made-in-USA strategy. Mirle plans to build autonomous quadruped robotics that mimic the movement of four-legged animals for industrial/factory environments. To enable this vision, Chelpis is exploring BrainChip’s advanced Akida™ IP to incorporate advanced visual GenAI capabilities in the proposed SoC design.

“The ability to add Edge AI security capabilities to our industrial robotics project that provides the low power data processing required is paramount to successfully achieving market validation in the robotics sector,” said Ming Chih, CEO of Chelpis. “We believe that BrainChip’s Akida is just the solution that we further need to bring our SoC to fruition. Their event-based processing and advanced models serve as a strong foundation for developing a platform for manufacturing customers looking to leverage advanced robotics in their facilities.”
“Akida’s ability to efficiently provide cyber-security acceleration with energy efficiency can help secure autonomous robotic devices,” said Sean Hehir, CEO of BrainChip.
“Akida’s innovative approach to supporting LLMs and GenAI algorithms could serve as a key contributor to Chelpis as they pursue government funding to develop their SoC and advance their industrial robotic initiatives.”

Actually it’s old news wrapped in new paper… it’s from April or mai …. or did i miss something on this „announcement“

 

[smoothsailing18]

World's first ?.

 

[7für7]

World's first ?.

It’s getting day by day better GO BRAINCHIP!!

 

[Pom down under]

That went straight on ignore.

Thought you meant me for a second, then I realized I wouldn’t be able to read this

 

[AusEire]

1. We all know the no 1 fan boy that made the company looked like it was about to go to the moon.
2. Look at previous AGMs where the CEO said there was going to be an explosion of revenue.

But yes, I made the decision to invest based on the hype.
Thanks for pointing that out.

Wait! So You've just admitted to the world that you invested in a company because essentially one person or in your words "fan boy" hyped it up online

And to your second point. Eh no. The CEO (LDN, Peter or Sean) at no stage DURING ANY AGM in the time I've been invested said that "there was going to be an explosion of revenue". Obviously they want revenue to increase but they never said that.

Given the conservative approach on announcing shit this simply never happened at an AGM. The only time that I can remember a statement being made with regards to sales in an AGM was by Antonio when he said that one contract could make Brainchip profitable over night and that was said at the 2024 AGM.

What you are referring to is an interview PVDM did with Commsec where he stated that he was expecting an explosion of sales to happen in time. That quote might not be exact but it's roughly what was said.

 

[Diogenese]

New GitHub update 20hrs ago on Akida/CNN2SNN and including TENNs release, modules and models etc by the looks.

@Diogenese will probs know more of anything unusual or new tucked in there.

Releases · Brainchip-Inc/akida_examples

Brainchip Akida Neuromorphic System-on-Chip examples and documentation. - Brainchip-Inc/akida_examples

github.com

20 hours ago

ktsiknos-brainchip
2.13.0-doc-1
d8435c2
Compare
Upgrade to Quantizeml 0.16.0, Akida/CNN2SNN 2.13.0 and Akida models 1.7.0
Latest

Update QuantizeML to version 0.16.0​

New features​

• Added a bunch of sanitizing steps targetting native hardware compatibility:
  • Handle first convolution that cannot be a split layer
  • Added support for "Add > ReLU > GAP" pattern
  • Added identity layers when no merge layers are present after skip connections
  • BatchNormalisation layers are now properly folded in ConvTranspose nodes
  • Added identity layers to enforce layers to have 2 outbounds only
  • Handled Concatenate node with a duplicated input
• Added support for TENNs ONNX models, which include sanitizing, converting to inference mode and quantizing
• Set explicit ONNXScript requirement to 0.2.5 to prevent later versions that use numpy 2.x

Bug fixes​

• Fixed an issue where calling sanitize twice (or sanitize then quantize) would lead to invalid ONNX graphs
• Fixed an issue where sanitizing could lead to invalid shapes for ONNX Matmul/GEMM quantization

Update Akida and CNN2SNN to version 2.13.0​

Aligned with FPGA-1679(2-nodes)/1678(6-nodes)​

New features​

• [cnn2snn] Updated requirement to QuantizeML 0.16.0
• [cnn2snn] Added support for ONNX QuantizedBufferTempConv and QuantizedDepthwiseBufferTempConv conversion to Akida
• [akida] Full support for TNP-B in hardware, including partial reconfiguration with a constraint that TNP-B cannot be the first layer of a pass
• [akida] Full support of Concatenate layers in hardware, feature set aligned on Add layers
• [akida] Prevented the mapping of models with both TNP-B and skip connections
• [akida] Renamed akida.NP.Mapping to akida.NP.Component
• [akida] Improved model summary for skip connections and TNP-B layers. The summary now shows the number of required SkipDMA channels and the number of components by type.
• [akida] Updated mapping details retrieval: model summary now contains information on external memory used. For that purpose, some C++/Python binding was updated and cleaned. The NP objects in the API have external members for memory.
• [akida] Renamed existing virtual devices and added SixNodesIPv2 and TwoNodesIPv2 devices
• [akida] Introduced create_device helper to build custom virtual devices
• [akida] Mesh now needs an IP version to be built
• [akida] Simplified model statistics API and enriched with inference and program clocks when available
• [akida] Dropped the deprecated evaluate_sparsity tool

Update Akida models to 1.7.0​

• Updated QuantizeML dependency to 0.16.0 and CNN2SNN to 2.13.0
• Sparsity tool name updated. Now returns python objects instead of simply displaying data and support models with skip connections
• Introduced tenn_spatiotemporal submodule that contains model definition and training pipelines for DVS128, EyeTracking and Jester TENNs models
• Added creation and training/evaluation CLI entry points for TENNs

Introducing TENNs modules 0.1.0​

• First release of the package that aims at providing modules for Branchip TENNs
• Contains blocks of layers for model definition: SpatialBlock, TemporalBlock, SpatioTemporalBlock that come with compatibility checks and custom padding for Akida
• The TemporalBlock can optionally be defined as a PleiadesLayer following https://arxiv.org/abs/2405.12179
• An export_to_onnx helper is provided for convenience

Documentation update​

• Added documentation for TENNs APIs, including tenns_modules package
• Introduced two spatiotemporal TENNs tutorials
• Updated model zoo page with mAP50, removed 'example' column and added TENNs
• Added automatic checks for broken external links and fixed a few
• Cosmetic changes: updated main logo and copyright to 2025

The reference to reLU is interesting:

"

• Added support for "Add > ReLU > GAP" pattern

"

There was a recent BRN talk (AL/JT ? maybe the Roadmap*?) which I understood to mean that ReLUs were replaced by LookUp Tables (LUTs) because ReLUs were very power hungry.


* Roadmap @ 7:30: https://brainchip.com/brainchip-technology-roadmap/ "activation function" in Akida 2
https://en.wikipedia.org/wiki/Activation_function

 

[Cardpro]

Wow 99%, only thing that's more accurate is our past guessing rate where we were proven (so far) to be nearly 100% wrong - Amazon's Alexa, Merc's MBOS, Valeo's Scala 3, Qualcomm snapdragon, Nintendo Switch 2, and other techs with Edge AI capabilities/low power consumption... lol

Hopefully we are finally in one!!!

Imo only dyor

An EXTRACT from this evenings press release.


View attachment 87322

Launch of BrainChip Developer Hub Accelerates Event-Based AI Innovation on Akida™ Platform with Release of MetaTF 2.13

BrainChip announces new Developer Hub and MetaTF toolkit, enabling seamless development and deployment of machine learning models on its Akida™ platform.

www.einpresswire.com

Hopefully...

View attachment 87323
View attachment 87324

 

[Fullmoonfever]

The reference to reLU is interesting:

"

• Added support for "Add > ReLU > GAP" pattern

"

There was a recent BRN talk (AL/JT ? maybe the Roadmap*?) which I understood to mean that ReLUs were replaced by LookUp Tables (LUTs) because ReLUs were very power hungry.


* Roadmap @ 7:30: https://brainchip.com/brainchip-technology-roadmap/ "activation function" in Akida 2
https://en.wikipedia.org/wiki/Activation_function

So given ReLU is power hungry & we replaced (as standard?) with LUT then is it possible someone specifically wants ReLU availability hence them still being listed in Git and docs?

 

[jrp173]

Wait! So You've just admitted to the world that you invested in a company because essentially one person or in your words "fan boy" hyped it up online

And to your second point. Eh no. The CEO (LDN, Peter or Sean) at no stage DURING ANY AGM in the time I've been invested said that "there was going to be an explosion of revenue". Obviously they want revenue to increase but they never said that.

Given the conservative approach on announcing shit this simply never happened at an AGM. The only time that I can remember a statement being made with regards to sales in an AGM was by Antonio when he said that one contract could make Brainchip profitable over night and that was said at the 2024 AGM.

What you are referring to is an interview PVDM did with Commsec where he stated that he was expecting an explosion of sales to happen in time. That quote might not be exact but it's roughly what was said.

 

[Doz]

The reference to reLU is interesting:

"

• Added support for "Add > ReLU > GAP" pattern

"

There was a recent BRN talk (AL/JT ? maybe the Roadmap*?) which I understood to mean that ReLUs were replaced by LookUp Tables (LUTs) because ReLUs were very power hungry.


* Roadmap @ 7:30: https://brainchip.com/brainchip-technology-roadmap/ "activation function" in Akida 2
https://en.wikipedia.org/wiki/Activation_function

ReLU strikes back …..

ReLU Strikes Back: Exploiting Activation Sparsity in Large Language...

Large Language Models (LLMs) with billions of parameters have drastically transformed AI applications. However, their demanding computation during inference has raised significant challenges for...

openreview.net

 

[Diogenese]

So given ReLU is power hungry & we replaced (as standard?) with LUT then is it possible someone specifically wants ReLU availability hence them still being listed in Git and docs?

The inner workings of Akida software are beyond my pay grade. However, the LUTs were introduced in Akida 2, probably as part of the 8 times efficiency improvement over Akida 1. If Akida 1 uses ReLU, this may be a tweak.

Further efficiency is provided in the secret sauce of the memory data movement patent applications which were in preparation at the time of the Roadmap broadcast.

My understanding* of ReLU activation function comes from the Wiki page. The neuron output is modified by multplying by the ReLU function. It could be a crude step function between -1 and +1, os the smooth S-shaped sigmoid transition, or some other comparable function:

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

* "understanding" is an exaggeration. "information about" is more accurate.

 

[AusEire]

You've gone a major league highlighting rampage.

Saying "They are looking forward to" is a bit different to "there will be an explosion of sales".

Saying that "there will be" implies that it will happen.