In some contexts, I'm pro-regulation... but I HATE how many countries (especially the USA) go about it.
My biggest pet peeve is prescriptive/closed-ended regulation, specifying EXACTLY what must be done... instead of specifying an outcome & allowing for multiple solutions.
For example: in the US, a lot of kids get run over by vehicles in reverse, due to blind spots.
Did they simply mandate that new vehicle designs eliminate those blind spots? NO, they SPECIFICALLY mandated back-up cameras 😬
The side effect of the back-up camera requirement is that EVERY new vehicle ends up with a fucking screen, a goddamn computer, and a shitload of electronics in it.
Now, I don't like cars, but I ESPECIALLY don't like enshittified computerized cars.
If they'd instead left it open-ended and mandated that blind spots be eliminated, we likely would've seen all sorts of innovative solutions... cameras, sure, but probably also some simple/reliable optical solutions involving mirrors and light pipes.
@IrelandTorin everything was headed in that direction anyway, starting with engines. Also, there's a lot more you can do with onboard computers that you can't do even with sophisticated optics, e.g., brighten an image.
@hallmarc Actually, you can brighten an image (and do quite a bit more image processing) without any onboard computers.
Ever heard of a photomultiplier tube or an image intensifier tube?
That's right, you can brighten an image purely in the analog domain, no computers required.
Believe it or not, photomultiplier tubes are actually WAY BETTER at capturing detail in low-light environments than traditional cameras with image brightening algorithms.They're even used in military night vision.
@IrelandTorin wouldn't they be unsuitable due to size/cost ratio and fragility? I mean I know that they can now be made much smaller than before but CMOS and CCD cameras are pretty well suited to the task. And yes, for cost purposes at scale, not to mention desirable features, upgradeability, and flexibility, a general computing device is going to almost always win out over analog devices or ASICS.
@IrelandTorin but the upfront investment for ASICs is bigger. Designing and fabricating them (and as you say getting it absolutely right) can take a lot longer. Scaling the design or upgrading takes longer. It's like building with concrete: rock solid and built to last but God forbid you need to change its design.
@IrelandTorin how do you upgrade a custom-built ASIC in millions of vehicles if something desperately needs to change?
@hallmarc Well, besides physically swapping out the ASIC, you don't.
But that is just as much a strength as a weakness - it essentially forces the manufacturer to do rigorous testing to verify that every aspect of the design works as intended before going to production... which in any safety-relevant application is a VERY good thing.
The software mindset is just bad. I say this as someone who once built software as a job.
It also ensures bad actors cannot, say, infect the system with a virus.
@IrelandTorin but designs change due to various factors, not just mission (or software) creep. Market pressures, new research, etc. In order to stay current/viable with ASICs, you'd need to over-engineer to 5, 10 years in the future. Meanwhile your competitor is pumping out designs that are easily modified and improved upon. I agree that safety designs should meet extremely high standards (look at aeronautics) and that simplicity should rule there but it always butts heads with profit.
@hallmarc If you're finding new ways to improve some metric of performance by altering the parameters of an existing engine via existing control pathways, you didn't do enough testing prior to developing your controls.
Aaaand taping out a new ASIC isn't THAT expensive. Given the scale automakers work at, & the degree of commonality between many vehicles of the same brand/family/year, they should have no problem pushing out a new ASIC every few years. Don't need a separate one for each model.
@hallmarc You may be surprised to learn that cost-at-scale often tilts strongly in favour of ASICs; the total silicon die area required for something like an engine ECU would be *tiny*, meaning your fab yields should be in the stratosphere, and those already super-high fab yields would be further enhanced by the use of more mature larger process nodes (unless things have changed, that probably means 28nm because that is - or at least was - the sweet spot for fab cost).