I see in the discussion that people seem to be very conservative or almost shy in regards to market share.
I see it a bit differently and would like to discuss this a little further, maybe I’m wrong?
As far as I can see, there are only 4 companies having anything commercially available that we can call neuromorphic.
A) Qualcomm
B) SynSense
C) GeneralVision
D)BrianChip
A) Qualcomm
Is the biggest question mark in my head, because I can’t get my head around exactly what it is that they have implemented? They were developing the Zeroth neuromorphic processor, but stopped the project and focused on a software stack instead. Now suddenly they have some similar abilities to Akida 1.0?
As far as I can see there are the following main scenarios:
1) They reopened the Zeroth project and finished the development of it.
2) They never gave up Zeroth, it was all deception and they just went into dark mode with it.
3) They used their experience with neuromorphic computing to quickly implement Brainchip IP.
So, this is in my view a flip coin that can land on the Qualcomm or Brainchip side.
They are huge in smart phones and could easily dominate edge AI in smartphones.
Besides that I don’t see them trying to sell IP, which means several things:
1) The design is not flexible to the purpose, which can be important in an era with a diversity of intelligent devices (IoT). An ADAS system might need more neuromorphic processing power and Nanose might be better off with a specialized power saving SoC.
2) Their chip has to exist besides what other chips are in the product, as it can’t be integrated. This in turn means bigger foot print, higher energy consumption and possibly slower data communication. They have very potent all-round chips, that can possibly do most of the tasks needed for many applications, but may not fulfill the low power requirements of smart home devices.
3) Competitors to Qualcomm will need to come up with a response, so they would like to integrate AI that is as/more potent and buy IP to integrate it.
B) SynSense
They have DYNAP-CNN that can be configured with up to 1 million neurons.
From what I can glean from it, they are focused on CNN models and doesn’t really support a wide range of neural networks like Akida (I think Akida supports vision transformers, RNN, DNN and SNN).
Also, it seems like they rely on back-propagation, to cite an article published on the SynSense website: “Some people approximate derivatives of spikes in order to use backpropagation (like SynSense) and some use another technique called spike timing dependent plasticity (STDP), which is closer to how biological brains function. STDP, however, is less mature as a technology (BrainChip uses this method for one–shot learning at the edge). There is also the possibility of taking deep learning CNNs (convolutional neural networks), trained by backpropagation in the normal way, and converting them to run in the spiking domain (another technique used by BrainChip).”
https://www.synsense.ai/synsense-cars-that-think-like-you-eetimes/
SyneSense does not seem to have any unsupervised learning capabilities (edge learning) or one-shot-learning like Akida.
SynSense does not seem to sell IP, but develops custom hardware. This probably means less drawbacks than Qualcomm has in this regard, but their developer team is going to be a bottleneck for wide adoption of their technology. Further considering their small size, it will be a drawback for them not to appear in eco systems like the Rensas, Megachips, ARM and Intel ecosystems.
Further SynSense may very well be hit by the US chip sanctions on China (the CHIP 4 act).
I don’t really see them capturing a big market share, but maybe I’m missing parts of the picture?
C) General Vision
In contrast to SynSense and Qualcomm, General Vision are selling IP.
The biggest neuromorphic chip I could find on their webpage has 5500 neurons…..
From my point of view I can only see General Vision capturing some very low cost focused and limited applications, maybe with volume, but not with a lot of revenue.
D) Brainchip
I like this half a year old chart that I think is made by Neuromorphia on this forum. It illustrates that we are already integrating widely and appeal to various use cases.
View attachment 32735
When looking at the pure plays Synsense, General Vision and Brainchip, we’re clearly the talk of the town, here from Google Trends:
View attachment 32736
Conclusion:
It’s hard to put a number on the future potential neuromorphic market share, but I think 10%-20% is underwhelmingly low considering the above. What do you think?