I've built simple CPUs out of logic gates and in simulators but the density of the circuitry in ICs, even old ones, never ceases to amaze me. To put things into perspective, that huge-looking wire coming onto the bond pad in the first picture is already thinner than a human hair. The smallest features in this 30-year-old CPU are a fraction of that size, and on a modern CPU are even smaller by a few orders of magnitude.
The true wonder there, I think, is the science and craft of photolithography.
Moore's law wouldn't have had the constant march forward it had for 40 years, if shrinking transistors smaller wasn't just the "simple" matter each time of either developing a slightly more accurate projection system; or discovering a quicker-reacting, less-runny, more-durable photoresist chemical.
We're running into the limits of this approach now, and having to look for other interesting tricks, but there was a lot of low-hanging fruit in photolithographic precision, there for us to pluck, and pluck, and pluck, year after year.
> but there was a lot of low-hanging fruit in photolithographic precision, there for us to pluck, and pluck, and pluck, year after year.
As i grow older i find myself recognizing similar trends in a number of fields. You have a initial slow takeoff period where basic science has to be figured out. Then you have a steep climb as that science unlocks a bunch of iterative refinements on the concept. And then you find yourself at a plateau as new iterations become expensive and complex.
I think this will hold true for the current interest around ML. We are already at a stage where we can not reliably 'debug' models.
And feeding more data in the pipeline seems to exhibit diminishing returns.
>"Moore's law wouldn't have had the constant march forward it had for 40 years, if shrinking transistors smaller wasn't just the "simple" matter each time of either developing a slightly more accurate projection system; or discovering a quicker-reacting, less-runny, more-durable photoresist chemical."
But isn't the "wall" we are hitting the result of quantum tunneling(leaking) and less about the photolithography process or am I misunderstanding your comment?
Its more of a combo of things at this point rather than specifically photolithography or unwanted quantum behavior that is slowing Moore's Law.
The currently used light wavelength is many times larger than the features they make with it, and the lamps for the next smaller wavelength have serious issues of not being bright enough as they are ~80w now and need to be ~200w, being blocked by air but they cant run in a vacuum because immersion lithography and other parts of the process don't work in vacuum and finally they are only reliable enough for lab use.
In other places they are running out of atoms to take away (eg: some parts are 15 atoms thick and at 14 atoms thick they no longer have the properties needed). In some cases they are stretching atoms farther apart to make them work better with the few they have (see strained silicon).
Another part is heat rejection from the silicon as while each transistor is more efficient and smaller on a individual basis, you have much more of them. This is why modern processors, when running a single thread workload, will shut off a core to spend its heat budget on running that 1 core faster or a big difference between a phone and tablet CPU is the amount of power dissipation rather than the silicon itself.
I wonder how much of the reverse engineering of these dies could be automated. Looking at these dies I think a lot of this could be done image mapping and then handing over annotated information of what pins go where to the person doing the re work. Might be pretty helpful to someone doing these sort of projects.
I'd also like to say kudos to the author. Great write up.
