cosors
👀
And Akida can still be stacked 64 times if I remember right, so I meant cluster. But that's certainly not what the 280 platform is for.Yes. The "P" would be used in servers where a lot of input signals need to be processed, possibly in the VVDN Edge Box, but "P" is very powerful - not really for battery powered devices. It can be up to 1300 times faster than "E".
I like the new ESP display with selectable specifications for each version:
https://brainchip.com/akida-generations/
Interesting to note that the memory per NPE increases to 100 KB in the "P", compared to 25 KB in the "S".
"S" (1 TOPS) is up to 10 times faster than "E", and "P" (131 TOPS) at max is 131 times faster than "S".
So a "P" NPE would have a larger footprint than an "S" NPE.
"E" (2 nodes) functions:
- vibration Detection
Anomaly Detection![]()
Keyword Spotting![]()
Sensor Fusion![]()
Low-Res Presence Detection![]()
Gesture Detection![]()
"S" (8 nodes) functions:
Advanced Keyword Spotting![]()
Sensor Fusion![]()
Low-Res Presence Detection![]()
Gesture Detection & Recognition![]()
Object Classification![]()
Biometric Recognition![]()
Advanced Speech Recognition![]()
Object Detection & Semantic Segmentation![]()
"P" (256 nodes) functions:
Gesture Detection*![]()
Object Classification![]()
Advanced Speech Recognition![]()
Object Detection & Semantic Segmentation![]()
Advanced Sequence Prediction![]()
Video Object Detection & Tracking![]()
Vision Transformer Networks![]()
* & recognition
A node has 4 NPEs. (neuromorphic processing engine)
I'm just waiting for someone to try the maximum possible even if it's only for research purposes. Maybe Brainchip could take this into their own hands like ARM does with their own chip to show what is possible.
It would be interesting to see the real-world comparison too which, as far as I know, we only know from a single AKD1000.
_____
8384TOPs could this be correct?
Even if there would be fewer TOPs in real terms as I read today.
https://www.eetimes.com/tops-the-truth-behind-a-deep-learning-lie/
Last edited: