Nvidia & Mercedes-Benz – The long cycle.

The 4-year design cycle problem again.

• Nvidia has landed a major blue-chip client for its Drive Orin platform in announcing with Mercedes-Benz but by the time it ships in 2024, the landscape may have changed highlighting again how problematic this 4-year design cycle is becoming.
• For those not familiar with the automotive industry, the vehicle design cycle is four years meaning that decisions for 2024 are being made now.
• This comes hot on the heels of the cancellation of Daimler’s autonomous driving partnership with BMW, clearly indicating that Mercedes-Benz wanted to go with Nvidia but BMW did not.
• BMW has historically been in the Intel/Mobileye camp which was strengthened when Intel purchased a 15% in the location technology company HERE in which BMW also owns a significant stake.
• Orin is Nvidia’s answer to the Tesla SOC that was announced at its 2019 analyst day that will be used to power all of the self-driving systems that Tesla’s will use in the future.
• Orin is more generic than Tesla meaning that it should be able to run autonomous driving solutions from a range of different providers giving Mercedes-Benz flexibility when the time comes to choose.
• This is imperative because the last thing it wants to do is get tied into its own solution (awful) or the solution of Nvidia (also awful).
• Instead, this should give it some flexibility to choose but this is where the 4-year design cycle starts to become a problem.
• Nvidia’s approach to AI-related problems is similar to that of Open AI (see here) which is to throw increasing amounts of compute power at difficult problems in the hope that deep learning will find a solution.
• This makes complete sense for Nvidia because the more compute power it sells, the more money it makes.
• However, I am not convinced that this is the right approach and that in four years’ time, the neural network heavy approach that Nvidia is taking may turn out to be suboptimal.
• If we ignore the issues with deep learning that make it totally unsuited for autonomous driving (see here), the verifiability issue alone is enough to raise serious questions especially in automotive.
• Vehicles need to be safe and in order to achieve an acceptable level of safety a vehicle system needs to be explainable such that when there is a problem, the reason for that problem can be found and dealt with.
• It is also important to be able to understand how a problem has been solved such that the user can be certain that the system will operate to the required level of safety.
• Deep learning and neural network-based systems are black boxes where the data goes in one end and the answer pops out of the other.
• There is very little, if any, verifiability meaning that achieving the desired level of safety for any system that is deep learning-based is going to be difficult.
• The greater the reliance on deep learning, the greater the difficulty and Orin appears going to be quite heavily reliant on deep learning systems.
• The direction the industry is now taking following the collapse of the deep learning hype in autonomous driving is more deterministic meaning rules-based software.
• This implies that the systems that will be available in four years time will be less reliant on deep learning meaning that the Nvidia architecture it has chosen for Orin may not be the right architecture for deployment in 2024.
• This could present Mercedes-Benz with some difficulties and again highlights that the 4-year design cycle for vehicles needs to begin excluding items where the technology is changing quickly.
• I have long believed that failure to fix this problem will leave the OEMs as also-rans in their own industry becoming vendors of smartphones on wheels.
• This is not a sustainable future as the OEMs have to find other revenue streams to offset the potential significant decline in vehicle demand caused by electrification and autonomous driving.
• Progress by the OEMs remains glacially slow.