Your 3.5GB is for a machine that runs a 32-bit operating system that knows only the most basic page-table algorithm. Such an operating system can form only 4GB of physical memory addresses, no matter how much RAM is fitted. Architectural decisions for the machine and peripherals mean that some of that 4GB is reserved for such things as a ROM BIOS, text-mode video memory, and a seemingly ever-increasing demand for memory-mapped devices, so that in practice only 3GB, give or take, of your potential 4GB is usable as general-purpose RAM.
However, a 32-bit operating system that knows the PAE page-table algorithm can form 64GB of physical memory addresses. On the same machine as above, the hardware architecture is unchanged. You still have that some of the first 4GB of physical memory is reserved, but memory above 4GB is usable by the PAE operating system as general-purpose RAM.
Microsoft's limit is that when you run 32-bit Windows with PAE, Windows may be made to ignore physical memory above 4GB. You get the same memory usage as if you did not have PAE.
After reading into the article, I see that you mention that 16bit Windows (limited by 64K) had 640K, so I do see that it has been (and still is) possible to go beyond the theoretical limitations of the OS archetecture.
The point there is that it's not a matter of going beyond the theoretical limitations but of realising that the theory was naive.
For DOS on the 8086, the mechanism by which 16-bit registers addressed a 20-bit physical memory space was primitive and ugly (as primitive things tend to be). Even programmers who didn't actually despise segmentation (though I suspect most did) had to code specially for it. Yet the mechanism can be seen as an early algorithm for translating linear addresses to physical addresses: instead of page tables maintained by the operating system, there were segment registers maintained by whatever was the currently executing code. Protected-mode segmentation was an improvement, but paging achieved elegance, especially because most programmers don't give it a moment's thought and many seem to be entirely ignorant of it. A down side to the way it was introduced is that many programmers got it stuck in their heads that linear and physical addresses must have the same width. A little history and reflection would have told them that the limitations they saw as derived from theory were actually signs of rigid assumption.
It really seems like it is many times easier to upgrade to x64 Windows.
It may or may not be. The article's concern is not with whether 32-bit Windows with PAE would be better or worse than 64-bit Windows, even on the narrow matter of using memory above 4GB. The concern is instead that you should know 32-bit Windows does have the capability coded into it as already sold to you, no matter how many people say that 64-bit Windows is the only way to use memory above 4GB.
Of course, some cases are obvious, e.g., if you were buying a new computer. Others I think are less so, or were. It is anyway all hypothetical since the license restriction means there is no alternative to 64-bit Windows, whether easier or harder. Without the restriction, I would have trouble advising someone to ditch a 32-bit installation that they're perfectly happy with except for not quite getting the use of all their RAM. A multi-GB download and upgrade or re-installation is not obviously easier than a license upgrade. If nothing else the latter involves a million times less data! Or would. It is, of course, not an option that Microsoft provides - or that anyone in the industry seems to have wanted that anyone in general should know is even possible.
I also have to add that the whole article really feels like an attack on Microsoft. While I agree with a lot of what you say about them being a monopoly and anti-competitive, etc, it seems like a little over the top for something that is fairly easily fixed.
If PAE support was not coded into the product, I would never have cared to ask for it. But since it is there, and Microsoft seems not to deal with it honestly, there seem to be some questions to raise.
for spreading that evil 3.5 meg limitation LIE.
