Great summary Dio,My understanding is that TENNs puts Akida in the shade for ultra-high speed, ultra low power. I'm not sure of the metrics for accuracy of inference. As for ML, TENNs uses "old school" back propagation, and it also uses MACs, although much smaller (ie, less power hungry) than conventional 16/32/64 bit MACs.
TENNs major claim to fame is it's ability to handle temporal information in silicon rather than software, which is a massive advance. On top of that TENNs can be advantageously implemented in software as part of the BRN software product line if silicon is not an immediate option.
I would expect that a TENNs licence would come at a significant cost premium over Akia1 because of its functional advantages, and since Akida 2 includes TENNs, that would cost more again. Akida 2 also adds long-skip which is important for example in NLP, and video processing. I don't know it there are any synergies between long skip and temporal processing, but they seem to me to be two sides of the same coin, and may be able to be combined advantageously.
We have also seen Akida 2 implemented in FPGE, albeit with inferior performance to what is expected of the Akida 2 ASIC. It had already been foreshadowed that the Akida 2 FPGA would be implemented as an on-line demonstrator.
Thinking about ML, Akida's feed forward (FF) ML would provide an enhancement to TENNs, and this may also apply to inference.
So the customer has a choice to trade off cost against performance. Where temporal analysis is crucial, TENNs would be the first choice. Akida 2 could add premium features such as advanced ML and object identification.
There is still the "stock-standard" Akida 1 in a variety of strengths which has many potential implementations, and I presume that it is available at a lower licence fee/royalty.
Then there is Pico, the Planck NPU, ideal for ultra low power applications. Interestingly, the original concept for Akida 1 was as a 1-bit NPU, but Pico could be implemented in up to 8 bits, giving much greater ability (*256) to discriminate between different input activations.
We don't know what future developments will produce, but BRN already has a broad spectrum of market-ready solutions. The company has not been sitting on its hands waiting for the market to catch up. It has laid a series of stepping stones for the customers to follow.
, cheers,Manny
No privacy risk
