BRN Discussion Ongoing

[Diogenese]

I usually post interesting (but also ridiculous) stuff from TSE in the German forum. This time however, the user “perhaps” who also is sometimes here active, postet today something interesting, I will share.

“I've taken a closer look at MYWAI because I see a lot of potential there. The MYWAI projects use a novel neuromorphic memristor technology (computing processes executed in memory) from General Vision.

If everything works as they claim, they have made a breakthrough in one of the biggest problems in neuromorphic computing. When the task changes, an automatic adjustment of the required neurons is necessary; otherwise, there will be increased inaccuracies. So far, there has been no solution to this problem. Akida is also affected by this issue. With General Vision's approach, it is theoretically possible to control the neurons of the Akida processor via the memristor, unlocking the full potential of Akida. Therefore, the progress of the MYWAI projects deserves increased attention, as this could be where the big breakthrough happens.

Sources: https://general-vision.com/download/neuromem-technology-reference-guide/?wpdmdl=12284&refresh=660275755e5f11711437173 https://neurotechnologijos.com/zusammenarbeit/?lang=de https://www.myw.ai/projects “

Not sure I understand the relevance of a problem with an analog neuron and Akida's digital architecture.

 

[7für7]

Not sure I understand the relevance of a problem with an analog neuron and Akida's digital architecture.

So far, the lack of an interface to implement the technology and make it mass-market compatible is the problem. That's why we are constantly hovering in the field of "testing" and quasi-successful integration of Akida into studies and prototypes, etc. It's about mass marketability.

 

[charles2]

View attachment 59896




Nvidia’s H100 AI #GPUs are taking the tech world by storm, but their reign comes at the price of a hefty energy bill.

According to a report from #CBInsights and #Stocklytics.com, these power-hungry processors are projected to consume a staggering 13,797 GWh in 2024, exceeding the annual energy consumption of nations like #Georgia and #CostaRica.

Imagine this, a legacy data center consumes 10 kW/rack where #CyrusOne, #KKR owned leading global data center operator and developer specialising in #AI applications, consumes 300 kW/rack!

But why do #GPUs consume so much power?

Data center #GPUs consume a substantial amount of power primarily due to their high computational requirements and the complex algorithms they handle. These #GPUs optimize parallel processing tasks like #machinelearning and #dataanalytics, involving simultaneous processing of vast amounts of data.

While #parallel processing speeds up data processing, one demerit is that, at a time most parts of a chip are active. This constant computation, coupled with the execution of complex algorithms, demands significant computational power, thereby increasing energy consumption.

The large-scale deployment of #GPUs in data centers, where racks and clusters utilize hundreds or thousands of #GPUs further amplify their collective power consumption. This combination of factors underscores the considerable energy consumption associated with data center GPUs.

Successfully navigating these challenges and fostering innovation will shape the future landscape of #AI computing.

So, what options do we have?

● #Amazon, frenemy to Nvidia, recently unveiled Arm based Graviton4 and Trainium2 chips holds promise for efficiency gains.

● In the near to medium term, #Neuromorphic computing is being researched aggressively as an alternative to synchronous parallel computing architectures. Neuromorphic computing is an asynchronous computing paradigm which runs on event based ‘spikes’ rather than a clock signal. And drastically lowers the power consumption.

● Big money is going into enabling tech like liquid cooling - #KKR acquired CoolIT Systems for $270 mn and Bosch acquired Jetcool through its venture arm

While CooIT Systems becomes the supplier for Cyrus One, #KKR makes money on both!

Never seen someone self describe as a polymath.

High bar but as the saying goes....if you got it, flaunt it.

Better be right though.

 

[SharesForBrekky]

As others have remarked earlier, Wavious seems to be out of business and their website is no longer accessible.

Apparently, their former CEO Benny Malek recently joined the Board of Endura Technologies:
View attachment 59908

Intriguingly, quite a number of previous Wavious employees (all with a Qualcomm background before joining Wavious, by the way, just like former CEO Benny Malek and former CTO Hanan Cohen) have since ended up at Apple - more than a coincidence? A quick Google search didn’t come up with any evidence of a takeover/acquisition of Wavious by the Cupertino-headquartered tech giant, though.


View attachment 59909

View attachment 59915
View attachment 59916



View attachment 59912
View attachment 59917

Good work Frangipani.

Here is an interesting website, which has some interesting connections:

Advancing semiconductor tech, transistor to connectivity

Semiconductor technologies are building blocks of modern tech. We enable advancements from transistor to connectivity via unique portfolio companies & platforms

www.bridgewestgroup.com

Bridgewest Group are very much involved with Endura Technologies as well as a couple of very interesting mentions with Wavious.

"Wavious offers a mix and match, plug and play chiplet platform that spurs innovation, accelerates time to market and reduces cost."

"2022
Semiconductor
Chiplet solutions company Wavious is acquired by a Fortune 5 company."

Couldn't find anything about this Fortune 5 company though.

 

[Diogenese]

I usually post interesting (but also ridiculous) stuff from TSE in the German forum. This time however, the user “perhaps” who also is sometimes here active, postet today something interesting, I will share.

“I've taken a closer look at MYWAI because I see a lot of potential there. The MYWAI projects use a novel neuromorphic memristor technology (computing processes executed in memory) from General Vision.

If everything works as they claim, they have made a breakthrough in one of the biggest problems in neuromorphic computing. When the task changes, an automatic adjustment of the required neurons is necessary; otherwise, there will be increased inaccuracies. So far, there has been no solution to this problem. Akida is also affected by this issue. With General Vision's approach, it is theoretically possible to control the neurons of the Akida processor via the memristor, unlocking the full potential of Akida. Therefore, the progress of the MYWAI projects deserves increased attention, as this could be where the big breakthrough happens.

Sources: https://general-vision.com/download/neuromem-technology-reference-guide/?wpdmdl=12284&refresh=660275755e5f11711437173 https://neurotechnologijos.com/zusammenarbeit/?lang=de https://www.myw.ai/projects “

Not finding any patents for General Vision (GV), but I found this one for Norlitech:

US2020082241A1 COGNITIVE STORAGE DEVICE 20180911

a system comprising a non-volatile storage memory, a controller, and a cognitive memory. The storage memory can store data. During operation, the controller programs a function for the system based on a configuration file. The function indicates one or more operations for the data stored in the storage memory. The cognitive memory can include a set of neuron memory cells, which can store a knowledge base for facilitating the function and execute a pattern matching operation between the data stored in the storage memory and the data stored in the set of neuron memory cells. The controller can then execute the one or more operations within the system based on an output of the pattern matching operation from the cognitive memory.

[0052] FIG. 1B illustrates an exemplary architecture of a CSD, in accordance with an embodiment of the present application. Search engine 130 can include a programmable hardware module 170 , which can be a configurable piece of hardware capable of accessing storage memory 150 , at least in part, to search for reference patterns loaded in cognitive memory 158 . Module 170 can execute firmware-level codes, and operate as a interface logic between storage memory 150 , cognitive memory 158 , cache 156 , and the host (i.e., storage node 116 ). In some embodiments, module 170 can be an FPGA-based module coupled to cognitive memory 158 . Module 170 can be based on one or more of: integrated circuitry, and a semiconductor intellectual property (IP) core. Cognitive memory 158 can include a neuron-based integrated circuit and/or a semiconductor IP core arranged as a single memory bank or a plurality of neuron banks 172 , 174 , and 176 . The components of cognitive memory 158 may be coupled via a PCB, or on multiple chips. Module 170 and cognitive memory 158 can also be parts of an integrated circuit on a common substrate (e.g., on the same die).

What makes this relevant to GV is the inventors are listed as GV employees:
PAILLET GUY; MENENDEZ ANNE

Looks pretty clunky.

 

[Fullmoonfever]

Not finding any patents for General Vision (GV), but I found this one for Norlitech:

US2020082241A1 COGNITIVE STORAGE DEVICE 20180911

View attachment 59920


a system comprising a non-volatile storage memory, a controller, and a cognitive memory. The storage memory can store data. During operation, the controller programs a function for the system based on a configuration file. The function indicates one or more operations for the data stored in the storage memory. The cognitive memory can include a set of neuron memory cells, which can store a knowledge base for facilitating the function and execute a pattern matching operation between the data stored in the storage memory and the data stored in the set of neuron memory cells. The controller can then execute the one or more operations within the system based on an output of the pattern matching operation from the cognitive memory.

[0052] FIG. 1B illustrates an exemplary architecture of a CSD, in accordance with an embodiment of the present application. Search engine 130 can include a programmable hardware module 170 , which can be a configurable piece of hardware capable of accessing storage memory 150 , at least in part, to search for reference patterns loaded in cognitive memory 158 . Module 170 can execute firmware-level codes, and operate as a interface logic between storage memory 150 , cognitive memory 158 , cache 156 , and the host (i.e., storage node 116 ). In some embodiments, module 170 can be an FPGA-based module coupled to cognitive memory 158 . Module 170 can be based on one or more of: integrated circuitry, and a semiconductor intellectual property (IP) core. Cognitive memory 158 can include a neuron-based integrated circuit and/or a semiconductor IP core arranged as a single memory bank or a plurality of neuron banks 172 , 174 , and 176 . The components of cognitive memory 158 may be coupled via a PCB, or on multiple chips. Module 170 and cognitive memory 158 can also be parts of an integrated circuit on a common substrate (e.g., on the same die).

What makes this relevant to GV is the inventors are listed as GV employees:
PAILLET GUY; MENENDEZ ANNE

Looks pretty clunky.

You might be interested in this GV presentation I just found which includes your Norlitech so good sleuthing D.

Full presso could maybe a few years old poss looking at the Rev dates?

It's HERE

 

[jtardif999]

A couple of years back Sean Hehir made a dsmissive statement that Akida 1000 should be seen as an early reference chip and the focus was going be on the later, more developed versions.

Since then Akida 1000, in chip form, has made it into Edge boxes, gone into space and looks like being our major income stream, It also goes out to universities, as chips alone or on sample boards. Good that we made a few chips up front.

Version 1500, taped out a while back, for 22nm, seems to have sunk without trace.

Version 2000 will not be made as a chip, probably because it is very expensive to make, perhaps beyond Brainchip’s resources. So we wait for potential users to be so attracted by its capability that they will fund the manufacture of their own chips, with our IP. And we wait and wait. Meanwhile the peloton closes in.

Not sure this all adds up to a convincing corporate strategy. There is a good case for physical chips to be available for early investigators and researchers, particularly if universities are seen as important cradles of understanding the full possibilities of AKIDA.

The Accenture patent is based on Akida2 IP imo. The patent describes prediction which is not an Akida1 capability.

 

[Fullmoonfever]

Not finding any patents for General Vision (GV), but I found this one for Norlitech:

US2020082241A1 COGNITIVE STORAGE DEVICE 20180911

View attachment 59920


a system comprising a non-volatile storage memory, a controller, and a cognitive memory. The storage memory can store data. During operation, the controller programs a function for the system based on a configuration file. The function indicates one or more operations for the data stored in the storage memory. The cognitive memory can include a set of neuron memory cells, which can store a knowledge base for facilitating the function and execute a pattern matching operation between the data stored in the storage memory and the data stored in the set of neuron memory cells. The controller can then execute the one or more operations within the system based on an output of the pattern matching operation from the cognitive memory.

[0052] FIG. 1B illustrates an exemplary architecture of a CSD, in accordance with an embodiment of the present application. Search engine 130 can include a programmable hardware module 170 , which can be a configurable piece of hardware capable of accessing storage memory 150 , at least in part, to search for reference patterns loaded in cognitive memory 158 . Module 170 can execute firmware-level codes, and operate as a interface logic between storage memory 150 , cognitive memory 158 , cache 156 , and the host (i.e., storage node 116 ). In some embodiments, module 170 can be an FPGA-based module coupled to cognitive memory 158 . Module 170 can be based on one or more of: integrated circuitry, and a semiconductor intellectual property (IP) core. Cognitive memory 158 can include a neuron-based integrated circuit and/or a semiconductor IP core arranged as a single memory bank or a plurality of neuron banks 172 , 174 , and 176 . The components of cognitive memory 158 may be coupled via a PCB, or on multiple chips. Module 170 and cognitive memory 158 can also be parts of an integrated circuit on a common substrate (e.g., on the same die).

What makes this relevant to GV is the inventors are listed as GV employees:
PAILLET GUY; MENENDEZ ANNE

Looks pretty clunky.

This is for the NeuroTile using the NM500. Again 2019.

HERE

 

[jtardif999]

I usually post interesting (but also ridiculous) stuff from TSE in the German forum. This time however, the user “perhaps” who also is sometimes here active, postet today something interesting, I will share.

“I've taken a closer look at MYWAI because I see a lot of potential there. The MYWAI projects use a novel neuromorphic memristor technology (computing processes executed in memory) from General Vision.

If everything works as they claim, they have made a breakthrough in one of the biggest problems in neuromorphic computing. When the task changes, an automatic adjustment of the required neurons is necessary; otherwise, there will be increased inaccuracies. So far, there has been no solution to this problem. Akida is also affected by this issue. With General Vision's approach, it is theoretically possible to control the neurons of the Akida processor via the memristor, unlocking the full potential of Akida. Therefore, the progress of the MYWAI projects deserves increased attention, as this could be where the big breakthrough happens.

Sources: https://general-vision.com/download/neuromem-technology-reference-guide/?wpdmdl=12284&refresh=660275755e5f11711437173 https://neurotechnologijos.com/zusammenarbeit/?lang=de https://www.myw.ai/projects “

“f everything works as they claim, they have made a breakthrough in one of the biggest problems in neuromorphic computing. When the task changes, an automatic adjustment of the required neurons is necessary; otherwise, there will be increased inaccuracies. So far, there has been no solution to this problem. Akida is also affected by this issue. With General Vision's approach, it is theoretically possible to control the neurons of the Akida processor via the memristor, unlocking the full potential of Akida. Therefore, the progress of the MYWAI projects deserves increased attention, as this could be where the big breakthrough happens.”

Where’s the proof that Akida is affected? You can’t go making statements like this without supporting evidence.

 

[7für7]

Not finding any patents for General Vision (GV), but I found this one for Norlitech:

US2020082241A1 COGNITIVE STORAGE DEVICE 20180911

View attachment 59920


a system comprising a non-volatile storage memory, a controller, and a cognitive memory. The storage memory can store data. During operation, the controller programs a function for the system based on a configuration file. The function indicates one or more operations for the data stored in the storage memory. The cognitive memory can include a set of neuron memory cells, which can store a knowledge base for facilitating the function and execute a pattern matching operation between the data stored in the storage memory and the data stored in the set of neuron memory cells. The controller can then execute the one or more operations within the system based on an output of the pattern matching operation from the cognitive memory.

[0052] FIG. 1B illustrates an exemplary architecture of a CSD, in accordance with an embodiment of the present application. Search engine 130 can include a programmable hardware module 170 , which can be a configurable piece of hardware capable of accessing storage memory 150 , at least in part, to search for reference patterns loaded in cognitive memory 158 . Module 170 can execute firmware-level codes, and operate as a interface logic between storage memory 150 , cognitive memory 158 , cache 156 , and the host (i.e., storage node 116 ). In some embodiments, module 170 can be an FPGA-based module coupled to cognitive memory 158 . Module 170 can be based on one or more of: integrated circuitry, and a semiconductor intellectual property (IP) core. Cognitive memory 158 can include a neuron-based integrated circuit and/or a semiconductor IP core arranged as a single memory bank or a plurality of neuron banks 172 , 174 , and 176 . The components of cognitive memory 158 may be coupled via a PCB, or on multiple chips. Module 170 and cognitive memory 158 can also be parts of an integrated circuit on a common substrate (e.g., on the same die).

What makes this relevant to GV is the inventors are listed as GV employees:
PAILLET GUY; MENENDEZ ANNE

Looks pretty clunky.

I’m not a technical expert but it’s better than the speculations about what happens to robs LinkedIn profile or if he flushed the toilet once or twice and if so, why he don’t like postings about Japanese toilets?

 

[7für7]

“f everything works as they claim, they have made a breakthrough in one of the biggest problems in neuromorphic computing. When the task changes, an automatic adjustment of the required neurons is necessary; otherwise, there will be increased inaccuracies. So far, there has been no solution to this problem. Akida is also affected by this issue. With General Vision's approach, it is theoretically possible to control the neurons of the Akida processor via the memristor, unlocking the full potential of Akida. Therefore, the progress of the MYWAI projects deserves increased attention, as this could be where the big breakthrough happens.”

Where’s the proof that Akida is affected? You can’t go making statements like this without supporting evidence.

It’s not my statement I just find it interesting and plausible.

 

[Fullmoonfever]

Not finding any patents for General Vision (GV), but I found this one for Norlitech:

US2020082241A1 COGNITIVE STORAGE DEVICE 20180911

View attachment 59920


a system comprising a non-volatile storage memory, a controller, and a cognitive memory. The storage memory can store data. During operation, the controller programs a function for the system based on a configuration file. The function indicates one or more operations for the data stored in the storage memory. The cognitive memory can include a set of neuron memory cells, which can store a knowledge base for facilitating the function and execute a pattern matching operation between the data stored in the storage memory and the data stored in the set of neuron memory cells. The controller can then execute the one or more operations within the system based on an output of the pattern matching operation from the cognitive memory.

[0052] FIG. 1B illustrates an exemplary architecture of a CSD, in accordance with an embodiment of the present application. Search engine 130 can include a programmable hardware module 170 , which can be a configurable piece of hardware capable of accessing storage memory 150 , at least in part, to search for reference patterns loaded in cognitive memory 158 . Module 170 can execute firmware-level codes, and operate as a interface logic between storage memory 150 , cognitive memory 158 , cache 156 , and the host (i.e., storage node 116 ). In some embodiments, module 170 can be an FPGA-based module coupled to cognitive memory 158 . Module 170 can be based on one or more of: integrated circuitry, and a semiconductor intellectual property (IP) core. Cognitive memory 158 can include a neuron-based integrated circuit and/or a semiconductor IP core arranged as a single memory bank or a plurality of neuron banks 172 , 174 , and 176 . The components of cognitive memory 158 may be coupled via a PCB, or on multiple chips. Module 170 and cognitive memory 158 can also be parts of an integrated circuit on a common substrate (e.g., on the same die).

What makes this relevant to GV is the inventors are listed as GV employees:
PAILLET GUY; MENENDEZ ANNE

Looks pretty clunky.

Excerpt from a 2021 Science Direct paper. NeuroEdge licenced GV NM500 and Neuromem.

https://www.sciencedirect.com/science/article/pii/S240595952100059X

3. System model and detailed description of the NeuroEdge computing system​

A detailed block diagram of the NeuroEdge system showing the role and placement of the NM500 chip, is presented in Fig. 1. The NeuroEdge system uses a neuromorphic device combined with Raspberry Pi to achieve real-time field training and inference on devices. NM500, however, is a neuromorphic chip whose neurons can learn and recognize patterns extracted from any data sources with less energy and complexity than modern microprocessors. NeuroEdge and NM500 are produced under the license of NeuroMem technology from General Vision Inc. [11]. Fig. 3 shows the structure of the NM500 and its application in various AI systems such as image, voice or video identification, classification, and anomaly detection.

 

[Diogenese]

So far, the lack of an interface to implement the technology and make it mass-market compatible is the problem. That's why we are constantly hovering in the field of "testing" and quasi-successful integration of Akida into studies and prototypes, etc. It's about mass marketability.

For the last couple of years, we have been focusing on the highly specialized market of IP licensing.

However the Akida AI Edge Box is directed to a much broader market, and it is pretty much plug-and-play. It's just a matter of loading the appropriate models and configuration.

 

[Fullmoonfever]

Not finding any patents for General Vision (GV), but I found this one for Norlitech:

US2020082241A1 COGNITIVE STORAGE DEVICE 20180911

View attachment 59920


a system comprising a non-volatile storage memory, a controller, and a cognitive memory. The storage memory can store data. During operation, the controller programs a function for the system based on a configuration file. The function indicates one or more operations for the data stored in the storage memory. The cognitive memory can include a set of neuron memory cells, which can store a knowledge base for facilitating the function and execute a pattern matching operation between the data stored in the storage memory and the data stored in the set of neuron memory cells. The controller can then execute the one or more operations within the system based on an output of the pattern matching operation from the cognitive memory.

[0052] FIG. 1B illustrates an exemplary architecture of a CSD, in accordance with an embodiment of the present application. Search engine 130 can include a programmable hardware module 170 , which can be a configurable piece of hardware capable of accessing storage memory 150 , at least in part, to search for reference patterns loaded in cognitive memory 158 . Module 170 can execute firmware-level codes, and operate as a interface logic between storage memory 150 , cognitive memory 158 , cache 156 , and the host (i.e., storage node 116 ). In some embodiments, module 170 can be an FPGA-based module coupled to cognitive memory 158 . Module 170 can be based on one or more of: integrated circuitry, and a semiconductor intellectual property (IP) core. Cognitive memory 158 can include a neuron-based integrated circuit and/or a semiconductor IP core arranged as a single memory bank or a plurality of neuron banks 172 , 174 , and 176 . The components of cognitive memory 158 may be coupled via a PCB, or on multiple chips. Module 170 and cognitive memory 158 can also be parts of an integrated circuit on a common substrate (e.g., on the same die).

What makes this relevant to GV is the inventors are listed as GV employees:
PAILLET GUY; MENENDEZ ANNE

Looks pretty clunky.

From last year's TinyML forum. We didn't appear to be in on this one.

tinyML On Device Learning Forum 2023

We held tinyML On Device Learning Forum 2023 on May 16th, 2023. To date, most ultra-low power machine learning (ML) applications at the edge are trained “off device” (typically in the cloud where virtually unlimited computing assets are available) while the edge devices perform the...

forums.tinyml.org

NeuroMem®, Ultra Low Power hardwired incremental learning and parallel pattern recognition
Guy Paillet, Co-founder and Chairman, General Vision Holdings

GV will present a Tiny RTML platform comprising of ST Nucleo64, together with a NeuroShield including 37 parallelized NM500 chips. This allows maintaining a parallel content addressable set of for example 21,000 Chinese characters. Submitting the image (16 x 16 pixels pattern) of a Chinese character, will return a category pointing on the English meaning within a constant search time of 30 microseconds. Learning time for additional character (on the spot learning) will also take about 30 microseconds per unknown character. The ANM5500 just released will make the same with only 4 chips and 5 times faster always, at milliwatts power. General Vision goal is to solve real world image recognition with learning and recognition on a small battery into for example a standalone (no network connection) Barbie doll, hence the patented “Monolithic Image Perception Device” successor of MTVS (Miniature Trainable Vision Sensor) allowing on “image sensor learning” and recognition.

Guy’s background is hardware design since 1976 starting with Motorola MC6800 as application engineer. He has been innovating on high performance Tiny Machine Learning since 1993 while inventing the ZISC36 with IBM Paris, Guy and family moved from France in 1996 and co-founder General Vision in 2000. Since, General has licensed its NeuroMem ZISC technology giving birth to 4 additional successful Neuromorphic AISC from 2007 to 2022, including the Intel Curie for “NeuroMEMS.”

 

[7für7]

For the last couple of years, we have been focusing on the highly specialized market of IP licensing.

However the Akida AI Edge Box is directed to a much broader market, and it is pretty much plug-and-play. It's just a matter of loading the appropriate models and configuration.

For which broader market do you think a costumer would take the AI edge box? This is again just a trial device in my opinion. More important is, what kind of steps our partner are doing in the ecosystem in my opinion.

 

[Diogenese]

This is for the NeuroTile using the NM500. Again 2019.

HERE

View attachment 59924

1152 silicon neurons - was this found at Cheops?

 

[Fullmoonfever]

1152 silicon neurons - was this found at Cheops?

Maybe....the Egyptians were pretty advanced a few thousand years ago.

 

[cosors]

For anyone speculative... answers in screenshot

Thanks Reuben for clearing that up for us, and thanks to you mrgds too!

TD I'm sorry, there are more important things to do. Please excuse us for keeping you from work today, we're just very interested.

 

[Justchilln]

A couple of years back Sean Hehir made a dsmissive statement that Akida 1000 should be seen as an early reference chip and the focus was going be on the later, more developed versions.

Since then Akida 1000, in chip form, has made it into Edge boxes, gone into space and looks like being our major income stream, It also goes out to universities, as chips alone or on sample boards. Good that we made a few chips up front.

Version 1500, taped out a while back, for 22nm, seems to have sunk without trace.

Version 2000 will not be made as a chip, probably because it is very expensive to make, perhaps beyond Brainchip’s resources. So we wait for potential users to be so attracted by its capability that they will fund the manufacture of their own chips, with our IP. And we wait and wait. Meanwhile the peloton closes in.

Not sure this all adds up to a convincing corporate strategy. There is a good case for physical chips to be available for early investigators and researchers, particularly if universities are seen as important cradles of understanding the full possibilities of AKIDA.

The Unigen cupcake uses akd1500

 

[chapman89]

Posted by Edge Impulse this morning.
There’s a lot of pages but I’ll just post some of it where it mentions Brainchip, as it won’t allow me to copy and paste for some reason, and those smarter than me can break down the whole thing if they have time and are willing.

https://assets.website-files.com/618cdeef45d18e4ef2fd85f3/65f826d1c27e243a45437112_Edge%20Impulse%20Industrial%20Productivity%20Solutions%20Guide%20v1.0_compressed%20(1).pdf?utm_content=287309981&utm_medium=social&utm_source=linkedin&hss_channel=lcp-19239958

What engine does Edge Impulse use to compile the Impulse?


It depends on the hardware.


For general-purpose MCUs, we typically use EON Compiler with TFLite Micro kernels (including hardware optimization, e.g. via CMSIS-NN, ESP-NN).


On Linux, if you run the Impulse on the CPU, we use TensorFlow Lite.


For accelerators, we use a wide variety of other runtimes, e.g., hardcoded network in silicon for Syntiant, custom SNN-based inference engine for Brainchip Akida, DRP-Al for Renesas RZV2L, etc.