More on Tesla Auto Pilot 2 (AP2) based on Elon's 2nQ Earnings call a few days ago, using tips from last Saturday's breakfast:
AP2 V9 SW coming "soon" a few weeks, will be big advance, but SW only.
AP2 Hardware announced during call, now in advance test, will ship in every car some time in 2019.
Fall 2017 Nvidia GPU upgraded in new cars since, with promise to retrofit all AP2 cars. This upgrade X10 performance, but now it has been supplanted by AP3. Elon promises it will be directly replaceable in existing AP2 cars.
Basic Neural Net operation is a convolution of a NxN subset pattern matrix against all the pixels in a scene, looking for close match to the pattern. This is a structured matrix multiply, highly parallel. GPUs can do this with far less fetches from main memory, but GPU is still too general purpose. Highest efficiency for this operation can be achieved with a Systolic Array, an old concept decades old, but only recently have chips had enough real estate for implementation. Basically the whole matrix operation is in hardware.
Google has had a Systolic Array chip they call TPU (Tensor Processing Unit) for at least a year, but they have not sold the chips, only access to TPU units in the cloud.
There have been reports that while AP2 has been using only its forward 3 cameras, the side cameras have been in use in "ghost" mode, in which deviations between how the driver is driving and how advanced AP would have driven are reported using data from these cameras.
Tesla calls there Systolic Array chip simply an "AI" chip. It should enable very good pattern recognition through all cameras. However, many important driving decisions require high-level rule-based logic. It is not clear AI chip could help this part.