Muniix wrote: ↑28 Jul 2018, 20:01
GrandAxe wrote: ↑28 Jul 2018, 19:33
Muniix wrote: ↑28 Jul 2018, 18:36
It DOES take huge amounts of compute to do ML/DL the phones you refer to have special TPU tensor processing units, the Huawei Mate 10 Pro, uses their own Kirin 970 that uses the GPU graphics processing units and the dedicated TPU cores for AI workloads. This Kirin 970 SoC is the same one I use for powertrain control atm, to do AI and physics compute loads.
The standard hardware the Formula One teams are forced to use have less than 1% of the compute of average mobile phones.
Ferrari may be putting in some intelligence into the TJI in engine, in that it's controlling it's own spark timing based on cylinder and pre-chamber MEMS pressure sensors, maybe ion sensing, and using some models with thermodynamics and detecting the reaction rate of radicals as they can be sensed this way with TJI and placing a 10mm spark plug just for ION sensing in main chamber, Ferrari could be a little bit more elegant in implementation.
1% the compute average of mobiles phones is pretty handy where the the dataset being processed is small and has few dimensions (unlike, say millions of dimensions in a jpeg or video) as will be the case with turning switches on or off based on a narrow set of parameters in an F1 car.
Machine learning algorithms are quite varied - from those that basically require more storage space and RAM than computational power and can run on bog standard devices (e.g. pos taggers); to systems that require super computers (e.g. weather forecasting and gene sequencing). Not everything requires GPU and other specialist hardware with frightening names.
Your reading sensor data, and processing it, it's not on/off single bit digital data, it's potentially more complex than recognising 'cats'.
Its surely not. Every sensor is a single dimension only.
A system as the one we are discussing might only want to know when the car is standing still, cornering, accelerating beyond a threshold, braking from above a threshold speed, one or two safety parameters; altogether only a handful of dimensions. Turning switches on or of in an F1 car should be pretty straightforward with the main challenges being integrating the machine learning system with the design/function of the car.
Recognising cats on the other hand, requires juggling millions of dimensions.
Talking about turning switches on or of in an F1 car to change functionality ... It could be anything from simple off/on switch actions, to loading custom software on the go or taking circuits offline to reprogramme them. The AI software could also be hidden like a virus to pop up when required. All these techniques can both change functionality of an engine and mask the method of change without requiring massive computational power.
I don't know, my interpretation is that it wouldn't be as a standalone unit. Meaning away from the existing MGU-H or turbo. I would say that in order to build up energy in the flywheel it would need to be recovered from somewhere else. I could be wrong, if you ask my wife I usually am! 
