It’s a fair comparison to look at what much more limited hardware could do decades ago, that far more advanced hardware can’t now. Especially when it appears to be an artificial software limitation.
We continue to disagree but I am fine with that.
There's a saying in software engineering that eventually all software will eat up all the resources provided by new hardware, which is why things don't run appreciably faster than they did 30 years ago. CPU's, memory, graphics all get faster, but software engineers load them down with so much new stuff, whether it be moving from 8-bit to 16-bit to 64-bit graphics and data or providing more and more sophisticated algorithms to handle bigger tasks.
I'll agree partially in that you can compare feature sets between hardware from 20 years ago to today's, but know that the amount of data being moved by today's computers is orders of magnitude larger than back then. And you have to compare apples to apples, not apples to oranges.
My disagreement is that we're talking about a feature on a desktop to a non-feature on a tablet, which is an apples to oranges comparison. It doesn't matter if the hardware was made the same day or 20 years ago. If the hardware isn't present to do something, it won't be able to do it well. Any desktop OS is built from the ground up to do virtual memory. All desktop CPU's and their chipsets are designed to handle that. The A-series was never built to do that. Optimized memory controllers are needed to handle moving large amounts of memory from RAM to disk and back while compressing/decompressing and cleaning up at lightning speeds. I used the ASICs comparison before where an ASIC built for the scrypt algorithm can easily mine Dogecoin because Dogecoin is built on that algorithm. But use a GPU, which is a general purpose unit, would take decades to mine a single coin where the ASIC would take weeks to do the same job. The ASIC is highly optimized for that one job and can do it far faster, just as an optimized memory controller is going to be able to manage virtual memory a lot faster than using non-optimized hardware. But that ASIC can't do much of anything else, which is why CPU's are composed of a mix of optimized hardware for specific tasks and general purpose hardware to run app code. When you're missing a part, it doesn't matter if some other hardware was able to do something 20 years ago.
None of that specialized hardware exists on the A12X or A12Z, so comparing something on a desktop of today to a desktop of 20 years ago is not meaningful. iPadOS is not a desktop OS, so its processors were never designed to handle desktop OS requirements until the M1 came around. Apple's workaround is simulating that dedicated hardware, likely with algorithms running on the high performance cores or GPU cores, none of which are optimized for handling virtual memory. They can do it, but not anywhere near as fast as dedicated hardware could. This is the same reason why a video editor can do H.264 rendering on any Mac, but a Mac with a media engine with dedicated hardware can do it immensely faster. That is the crux of the issue. As Craig F said months ago, they ran SM on many iPads but found the performance inadequate on non-M1 iPads. That remains true. But they were able to find a workaround that brought a limited feature set to the older iPads. That workaround is not going to be able to do the job of a dedicated "engine" nearly as fast or efficiently.
On top of that, the older iPads have less memory and flash that's five times slower than the flash on the 2021 M1 iPads compounding the difficulty. Consider people were complaining about slow flash on the base M2 MBA that was half the speed of the M1 MBA and complaining that the performance was unacceptable. The flash on the 2018/2020 iPads are 1/5 the speed of the 2021 M1 iPads. That comparison should put some things in perspective.
