BRN Discussion Ongoing
- Thread starter TechGirl
- Start date
Now edited thanks Dodgy Knees! 
Untether AI Touts “At-Memory” Architecture, Promising Efficiency And Performance At The Edge
What does Untether AI promise?
Untether AI recently emerged from stealth and described their architecture as being “At-Memory”, which the company claims will deliver leading performance and power efficiency.So, let’s look at the architecture. I like the chart below, as it clearly positions Untether AI versus cache-oriented Von Neumann designs. While interesting, the assertion that this architecture is unique is a bit of a reach: other startups have somewhat similar memory architectures, but few can claim the efficiency that Untether AI is demonstrating with early silicon.
Now, here’s the thing: Untether AI claims that its runAI200 chip will crank out over 500 Trillion Ops Per Second (TOPS) with 8-bit inference operations with 200MB of on-die SRAM, all at 8 TOPS/watt. That implies the chip consumes about 60 watts or so, while potentially cranking out 80% of the inference performance of the industry-leading NVIDIA A100, which consumes 300-500 watts. ( IS 60W TOO MUCH FOR AKIDA ??? )
For applications such as ADAS and mobile autonomy, latency can be as critical as throughput, and Untether seems to shine here. The company is shipping chips, cards, and software by the end of this year to early potential clients.
Conclusions
While a few competitors will claim better power efficiency, none that we are aware of can also claim exceptional inference performance and latency. UntetherAI may be setting a new bar here, and may have significant opportunities in performance-demanding applications such as self-driving vehicles.
Untether AI Touts “At-Memory” Architecture, Promising Efficiency And Performance At The Edge
The startup’s unique architecture looks promising (but so do all Startup’s)!
www.forbes.com
VW had been collaborating with Audi and Porsche on LiDaR/LSTM/SpikingNN since at least 2017.
US10133944B2 Digital neuromorphic (NM) sensor array, detector, engine and methodologies
US10387741B2 Digital neuromorphic (NM) sensor array, detector, engine and methodologies
[042] ... The digital NM vision system also can include a digital NM engine 145 that perform image and data processing operations on the velocity vector data generated by the digital NM detector that enables image data processing for improved object detection, classification, and tracking, including machine and deep learning. As such, in accordance with at least one embodiment, the digital NM engine 145 may include one or processors running software to generate digital NM output data for analysis and subsequent control of components with the environment imaged by the detector 110 . Operation of the digital NM engine 145 is further discussed herein with connection to FIGS. 17-20.
That said, more efficient ways of processing the the data than "processors running software" are now available.
However, Akida 1000 does not do tracking, so they will need to wait for Akida 2000 which will include LSTM, and is being developed at top speed by BRN.
LSTM is probably available in MetaTF/ADE software for simulation testing as we speak.
US10133944B2 Digital neuromorphic (NM) sensor array, detector, engine and methodologies
US10387741B2 Digital neuromorphic (NM) sensor array, detector, engine and methodologies
[042] ... The digital NM vision system also can include a digital NM engine 145 that perform image and data processing operations on the velocity vector data generated by the digital NM detector that enables image data processing for improved object detection, classification, and tracking, including machine and deep learning. As such, in accordance with at least one embodiment, the digital NM engine 145 may include one or processors running software to generate digital NM output data for analysis and subsequent control of components with the environment imaged by the detector 110 . Operation of the digital NM engine 145 is further discussed herein with connection to FIGS. 17-20.
That said, more efficient ways of processing the the data than "processors running software" are now available.
However, Akida 1000 does not do tracking, so they will need to wait for Akida 2000 which will include LSTM, and is being developed at top speed by BRN.
LSTM is probably available in MetaTF/ADE software for simulation testing as we speak.
DingoBorat
Slim
Whether 70 or 300 sensors, on a car, wouldn't/couldn't they be served by just a few AKIDAD chips?@chapman89 my calculator broke just on that many sensors let alone my convo with PVDM chatting about their are up to 300 sensors on a automobile , so if we get a number between 70 and 300 , I'm gonna need another wider calculator
So there may only be 10 or even fewer royalties per car..
With 300 sensors, there wouldn't be 300 "brains"..
We have one brain, for the five senses we use..
DingoBorat
Slim
Isn't "deep learning" purely cloud based and the complete opposite of what we do Diogenese?
DingoBorat
Slim
"That implies the chip consumes about 60 watts or so, while potentially cranking out 80% of the inference performance of the industry-leading NVIDIA A100, which consumes 300-500 watts"Untether AI Touts “At-Memory” Architecture, Promising Efficiency And Performance At The Edge
What does Untether AI promise?
Untether AI recently emerged from stealth and described their architecture as being “At-Memory”, which the company claims will deliver leading performance and power efficiency.
So, let’s look at the architecture. I like the chart below, as it clearly positions Untether AI versus cache-oriented Von Neumann designs. While interesting, the assertion that this architecture is unique is a bit of a reach: other startups have somewhat similar memory architectures, but few can claim the efficiency that Untether AI is demonstrating with early silicon.
Now, here’s the thing: Untether AI claims that its runAI200 chip will crank out over 500 Trillion Ops Per Second (TOPS) with 8-bit inference operations with 200MB of on-die SRAM, all at 8 TOPS/watt. That implies the chip consumes about 60 watts or so, while potentially cranking out 80% of the inference performance of the industry-leading NVIDIA A100, which consumes 300-500 watts. ( IS 60W TOO MUCH FOR AKIDA ??? )
For applications such as ADAS and mobile autonomy, latency can be as critical as throughput, and Untether seems to shine here. The company is shipping chips, cards, and software by the end of this year to early potential clients.
Conclusions
While a few competitors will claim better power efficiency, none that we are aware of can also claim exceptional inference performance and latency. UntetherAI may be setting a new bar here, and may have significant opportunities in performance-demanding applications such as self-driving vehicles.
Untether AI Touts “At-Memory” Architecture, Promising Efficiency And Performance At The Edge
The startup’s unique architecture looks promising (but so do all Startup’s)!
I think... Maybe they should have come out of stealth 5 years ago, when 60 watts power usage, would have been revolutionary
Last edited:
Yes thanks…I thought that might of been too much"That implies the chip consumes about 60 watts or so, while potentially cranking out 80% of the inference performance of the industry-leading NVIDIA A100, which consumes 300-500 watts"
I think... Maybe they should have come out of stealth 5 years ago, when 60 watts power usuage, would have been revolutionary
If the Akidas are integrated with sensors, there would only be one per sensor, but it would only be a partial Akida, eg, 4 nodes or so instead of 20. Then there's mission-critical function redundancy. I recall in the early days of electronic ignition, Holden had a "limp-home" function (30 kph), which, on one occasion, came into play returning to Sydney from Wollongong Uni.
With any luck there will be at least 20 full Akida equivalents at, say, $10 per pop. After all, we're not here to strangle the golden-egg-laying goose.
DingoBorat
Slim
Bravo
If ARM was an arm, BRN would be its biceps💪!
Check this out team!
I thought these slides from NVDIA's Investor Day 2021 Presentation were pretty interesting. The first slide shows the NVDIA + ARM relationship and the second shows a Business Model with Mercedes for shared revenue for AutoPilot and AI cockpit software per car, which I imagine is like Netflix-style subscription service to enable features like autonomous driving, from which NVDIA will presumably take some kind of cut???
I thought these slides from NVDIA's Investor Day 2021 Presentation were pretty interesting. The first slide shows the NVDIA + ARM relationship and the second shows a Business Model with Mercedes for shared revenue for AutoPilot and AI cockpit software per car, which I imagine is like Netflix-style subscription service to enable features like autonomous driving, from which NVDIA will presumably take some kind of cut???
Hi DB,
"deep learning" refers to learning on a multi layer NN such as a CNN.
With CNN2SNN, Akida has been adapted to work with deep learning trained CNNs.
https://tech-writing-graphics.com/wp-content/uploads/2019/06/Akida-CNN2SNN-toolKit-v.1.1.pdf
Boab
I wish I could paint like Vincent
Haha, Napoleon Dynamite always makes you feel good.
Dozzaman1977
Regular
Love this Quote from Rob telson about the new podcast guest...... ... ...
"Jan is one of the people who we work with to help drive neuromorphic compute and ML at the Edge to enable more intelligence for billions of IoT devices," said Telson.
BILLIONS BILLIONS BILLIONS
"Jan is one of the people who we work with to help drive neuromorphic compute and ML at the Edge to enable more intelligence for billions of IoT devices," said Telson.
BILLIONS BILLIONS BILLIONS
DingoBorat
Slim
hamilton66
Regular
K, ur right - 2022 is more than halfway down. As we sit right now, BRN is up over 50% from 1/1/2022. People are looking at perceived losses from highs, rather than looking at historical gains. I want it all, and I want it now. Those with a l/t strategy need to chill. We're going great. Plenty of positive news flow to come over the coming 6 mths IMO. Patience. GLTA
Is it just me, or everytime I read the word compute I have to say it in an Anastasia type accent?Love this Quote from Rob telson about the new podcast guest...... ... ...
"Jan is one of the people who we work with to help drive neuromorphic compute and ML at the Edge to enable more intelligence for billions of IoT devices," said Telson.
BILLIONS BILLIONS BILLIONS
Attachments
RobjHunt
Regular
S
Straw
Guest
Well her sister-in-law (or equivalent) certainly won't have to explain to her why Brainchip is a good investment!
Deadpool
hyper-efficient Ai
She does it for me
uiux
Regular
Conversation over beers last night with brother in law (however could have been anyone near and dear).
Him: “eff sake didn’t win lotto again…”
Me: “mate you’d be better off to put this in stock…”
Him (laughs): “yeah right, like you are some stock portfolio mogul…”
Me: “Nope, but I do have a good pile in Brainchip and frankly this very much has potential to be life changing for us”.
Him: “what, you now gambling like bitcoin…”
I then proceeded to try and firstly outline Akida and then mention all the partnerships and announcements we’ve had. I could see him glaze over and thinking I was on crack when talking about the huge potential for this to take flight. And then he started his usual way of ridiculing people (very much a defense mechanism when he doesn’t understand
I however also realised, there was no page or post which just shows potential use cases and partnerships etc etc. kinda a one pager listing all we know so far to show to people.
It would be cool to have such a document to show the skeptics, or does this already exist?
Cheers and happy Sunday!
I've spent 7+ years researching the fuck out of the neuromorphic industry and this is the one pager I have produced showing the potential use cases for neuromorphic technologies :
Fundamental Analysis
HI all. I have been sitting back just watching as the chaos settles thinking of how to make this site a more friendly place for fundamental analysis and also a more easily readable and navigable resource. As everyone is aware I love contributing research that has been collated and themed...
thestockexchange.com.au
Similar threads
