Given the BRN/Valeo JD partnership, and the fact that Akida 2 has not made it to silicon, I've been developing the hypothesis that Valeo Scala 3 is provided with Akida software for classification of point cloud objects, so I looked at some recent Valeo patents which do use CNN software running on a system computer for point cloud classification. Using software would have made it possible to continually upgrade the system as TeNNs/ViT was refined.
"
At Valeo, we're harnessing AI and GenAI to revolutionize the automotive industry, from enhancing EV range and smart lighting to advancing autonomous driving,” Merlin shared. “We’re integrating AI into every aspect of our business, ensuring rapid development and heightened security."
This Valeo patent relates to training a NN:
US2023146935A1 CONTENT CAPTURE OF AN ENVIRONMENT OF A VEHICLE USING A PRIORI CONFIDENCE LEVELS - 20211109
[0014] T
he steps of capturing the environment of the vehicle by the at least one environment sensor and of processing the point cloud using the trained artificial intelligence for the content capture of the environment relate to the operation of the vehicle. These steps are therefore each performed individually in each driving assistance system. These steps are furthermore performed repeatedly in the driving assistance system in order to perform continuous content capture of the environment.
In effect, this would relate to the preparation of NN models from lidar data.
This application was filed in October 2022:
WO2024088937A1 METHOD TO ANALYZE AT LEAST ONE OBJECT IN AN ENVIRONMENT OF A SENSOR DEVICE FOR A VEHICLE 20221027
T
he analysis algorithm may, in particular, comprise a convolutional neural network. The analysis algorithm is at least configured to, for example, detect, identify and/or classify the object. By means of the convolutional neural network, it is, for example, possible to perform a classification of the object if the convolutional neural network is trained to classify objects.
It seems highly probable that BRN were working with Valeo to develop NN software to process lidar signals.
Remembering also that the TeNNs patent applications were filed in June 2022:
WO2023250093A1 METHOD AND SYSTEM FOR IMPLEMENTING TEMPORAL CONVOLUTION IN SPATIOTEMPORAL NEURAL NETWORKS 20220622
it would not have been possible to have TeNNs silicon prepared at that time. In fact we know that TeNNs was not taped out until much later.
This being the case, and in view of the recent Valeo patents I'm prepared to bet that the BRN/JD partnership encompasses the development of software NN, including the creation of NN models. The milestone payments would have included model development milestones.
So, in my view, there is no Akida SoC in Scala 3, but there is Akida NN model software provided with Scala 3 to run on the ADAS processor, and adaptable to run on an Akida 2 processor with TeNNs when the silicon is available.
I can't say whether or not this is the model that Mercedes has adopted, but since they don't have Akida 2 silicon, there is a high probability that it is the case.