>>14403678Because you need a use case to make a chip for, and finding the use case that justifies the cost you'll put into making a chip needs to be worth it, because that use case is something you should be able to accelerate and then make money from to make back the cost. Additionally, once you have made the chip, you need to write the drivers for it. Once you've written the drivers, you then need to write an SDK for it. Once you have an SDK, you need to write your first useful program for it.
It's not that the chips are hard to make, its that every step that comes after increases by 10x in complexity. This means that if it costs you 100 to make the chip, it will cost you 1000 to make a good SDK, and 10,000 to make a good program that shows off the capability of the chip.
Take Tesla's FSD computer. Its 2 DNN chips running in parallel, with 1 in shadow mode and one in active/production mode. Tesla was relying on Nvidia's driveGX or whatever for the better part of a decade but then decided to build their own ASIC when they realized that the GPU wasn't moving fast enough and couldn't process images fast enough. So they wrote their own silicon for it, which gave them a 5-7x improvement over any hardware Nvidia puts out at 1/2 the power use. But once they had the silicon, they had to rewrite their entire software stack to take advantage of it. And now, after doing that for the better part of the last 5 years that way, they're writing their first good program that shows off the capability of that chip. How? By making the Dojo supercomputer which is essentially the FSD DNN but evolved and at megascale so that it can do ExaFLOP FP16/32 scale training and drop Tesla's training time from ~1 week to ~1 day.
So then for ^ consider just how much money Tesla has spent for designing their own ship, rewriting their entire stack, and then creating Dojo (as the program to show it off to the world). Hence rich companies/countries don't just go rogue on chips. You need a purpose.
