Could it be that Mobileye is overrated? After the astounding $15 billion acquisition by Intel, many assumed that the way to a fully autonomous car must pass through a hardwired chip with hardcoded software. That was the Mobileye way, and it indeed seemed valid at the beginning of the computer vision revolution.
However, AI has evolved tremendously since those early days to a point that could soon render the old paradigms irrelevant. Most of that evolution revolves around moving from classical computer vision to more flexible deep-learning algorithms, which don’t necessarily require massive dedicated hardware anymore. These software-based solutions can run on a variety of processors and the one offered by Brodmann17, for example, can even run on low-power CPUs.
Despite that, Mobileye still offers somewhat of a cookie-cutter solution for a market with ever-changing requirements. This troubles quite a few Mobileye customers, as when your supplier only sells hammers, your problem better look like a nail. Perhaps all automotive-related challenges look like nails then? It is enough to spend a day with any leading automaker to realize that this is wishful thinking at best.
In contrast, the deep-learning algorithms offered by Brodmann17 for example, can run on any processor, and work with any standard framework and any runtime library. Forget about cookie cutters, this flexible technology can be easily adapted to provide a tailor-made solution for a variety of needs, from handling computation of in-cabin and reverse cameras to front camera processing of autonomous cars. As it runs 20X faster than standard deep-learning algorithms, it can boost AI performance beyond all expectations when running on any given hardware.
The unprecedented efficiency and accuracy of such algorithms have closed the gap between the past compulsion to create rigid hardware solutions and the present need for flexibility. These hardware-agnostic algorithms have leapfrogged hardwired solutions in all major aspects – performance, accuracy, flexibility, and pricing. There is no more need for a hardware overkill with its heavy toll of rigidity, as even low power processors of the pre-AI era can efficiently run those pioneering algorithms. A software-based solution can therefore offer different levels of compute power and price points to perfectly match each use case, and help automakers turn sci-fi into reality a lot sooner.