Monkeymandan
Regular
Very well said. That’s how to welcome a new holder - as I am.Hi Semmel,
Welcome to the BRN discussion. It will make for a change from reindeer politics.
Your timing is impeccable as I doubt the price will ever be this low again.
The BRN business model is licensing the IP for the Akida digital SNN processor IP, in a similar manner to the way ARM licences its RISC-IV processor IP. There is a "sign-on" licence fee of a few million $AU, and royalties are paid on sales.
Akida consists of an array of neuromorphic processing units (NPUs). The NPUs are grouped into nodes (4 NPUs per node). The number of nodes can be chosen depending on the complexity of the intended task from 2 nodes up to 64 nodes. Several Akida chips can be combined for even larger tasks. Akida 2 will come in 3 different classes, E being 2 to 4 nodes, S being up to 8 nodes, and P being up to 64 nodes. The licence fee and royalties will be calculated on the number of nodes chosen.
Renesas has a licence for 2 nodes which will be integrated into a microprocessor, and this is said to be sufficient to perform inference (object classification/identification) at 30 frames per second.
Megachips has a licence for Akida although we have not seen the details we know that they were involved in the development of Akida 1500, which involved the production of the Akida chip by Global Foundries on 22 nm FDSoI (Fully depleted Silicon on Insulator, a very low leakage loss technology.
Akida has been produced by Vorago in radiation hardened format for NASA.
Mercedes used Akida in its EQXX giving the example of the key word spotting being 5 to 10 times more energy efficient than alternatives.
BRN has been working with the lidar manufactures Valeo for some years, and Akida is said to have a sweet spot for lidar.
Akida has also been trialed with Prophesee's DVS (dynamic vision system) event camera performing object classification such as hand gesture recognition, again greatly outperforming the Synsense model.
Akida is processor-agnostic and can be used with any sensor input. It is "programmed" by an associated CPU processor loading model libraries and weights and configured into the appropriate arrangement of layers and nodes, but, once configured, the CPU does not take part in the inference/classification which is all carried out in the Akida silicon without any software being involved. This makes Akida both highly energy efficient and extremely high speed.
Akida also uses N-of-M coding, which is much more efficient than rate coding previously used in NNs. This coding is based on the fact the information is carried in the first spike to arrive from an optical sensor, while rate coding counts the spike frequency to extract the information. BRN purchased Spikenet (Simon Thorpe's spin-off company) which developed N-of-M coding.
This is the patent for the NPU along with my favourite drawing:
US11468299B2
View attachment 46151
Footnote: For those who think @Semmel is a troll or is lazy, I recommend you have a look at the contributions to the Talga threads. Semmel's contributions and depth of analysis are a major cornerstone of the complex discussion of the native title court issues Talga is subjected to.
Trawling through this thread BRN would be a very difficult and inefficient way of gathering information for someone starting from scratch.
Thanks to you and @cosors for unknowingly pricking my interest in BRN a few months back.