butcherano
Regular
Just one more piece of ammunition for Akida being embedded on the Renesas ZMOD gas sensors is the improvement that they’ve made to the power consumption over the last couple of years.Hi @butcherano ,
Just between you and me, I think there is a possibility that Akida could be used in these sensors, but keep it under your hat for now.
This Renesas patent has an early priority date of 20191323, and a second priority date of 20201215. That may indicate that some new tech was included in the second priority document. It would be nice to find out what new tech was disclosed in the second document.
The specification does discuss NNs in a general manner, but does not disclose the nitty-gritty of how to make a NN circuit, so it would be interesting to know where they got their NN from, n'est-ce pas? I suppose we could count all the available NN SoC IPs available at the time.
They have also thrown in a reference to analog NNs to cover themselves.
It's also interesting to compare the patent drawings with the one in your abstract.
View attachment 10494
US2021190750A1 System and Method to Avoid the Influence of Ozone for a Gas Sensor
Priority: 20191223; 20201215.
Note 2 priority dates indicates new material was added in filing the complete application US202017122761A.
View attachment 10492
View attachment 10493
As illustrated in FIG. 1C, AI processing may be a neural network 142 . The main components of neural network 142 are input nodes 148 that receives the data from pre-processing 140 , nodes in hidden layers 146 where data is processed by sequentially applying weighted functions to process data and pass to the next layer. Finally, output nodes 150 provided the results. The weighted parameters of each of the functions executed in each of the nodes can be set by a training algorithm that rely on a training data set. Memory 124 is large enough to store the weighting parameters for operation of neural network 142 . In some cases, neural network 142 may be, at least partially, trained in a supervised learning method where gas sensor 100 processes a set of known gas samples to set the weighting parameters. In either case, neural network 142 may provide much more reliable results for the gas dependent results output by gas sensor 100 .
[0041] Neural network 142 can be implemented digitally in processor 122 or may be implemented in analog circuits in processor 122 . In some embodiments the weighting factors to each node may be digitized and the nodes of neural network 142 may be implemented by analog circuitry. Whether processing in each node is performed digitally or by analog circuits, the weighting factors are stored in memory 124 and coupled to each of the nodes of input nodes 148 , hidden layers 146 , and output nodes 150 .
As I said, don't tell the others - we don't want to start a riot - at least til I've got some more at these prices.
From the 1st generation back in 2020 to the 2nd generation available now, the power consumption has improved by 3 orders of magnitude from 1.5mW to only 160uw.
Akida would definitely be capable of achieving this. I can’t see any other changes that might account for this massive improvement.
June 2020 datasheet: https://docs.rs-online.com/e05c/A700000007499355.pdf
Dec 2021 datasheet: https://www.renesas.com/us/en/document/dst/zmod4410-datasheet
Last edited: