Originally Posted by Martinus2u
it is relevant in so far as it limits the amount of physical memory the system can use. (note we are not talking about virtual address space here, nor demand paging.)
Actually the 800MB (approx.) is the limit on kernel virtual memory space. The limit is the 1GB of process virtual space reserved for the kernel minus the space required to map PCI bus addresses. It does not limit the amount of physical RAM that the kernel or applications can use. The 800MB limits how much physical RAM the kernel can leave permanently mapped for access by the kernel. The kernel doesn't permanently map all of physical RAM. Instead the kernel temporarily maps individual pages or small ranges of pages when it needs to access the physical RAM. The limited kernel virtual space may affect performance but it doesn't limit the amount of physical RAM.
The actual physical RAM limit for a 32-bit kernel is 64GB. That is large enough that very few systems would ever encounter the limit. Note that this is not true of Windows, where Microsoft purposely limited non-server 32-bit versions of Windows to 4GB of RAM (by leaving out PAE support).
In fact the physical RAM on most systems is limited by the chip-set rather than 32-bit software. Intel made a marketing decision to limit their non-server chip-sets to 4GB until recently. Intel, like Microsoft wanted to charge a premium for its "server" chip-sets that supported over 4GB of RAM.
In some respects Intel and Microsoft were both dragged "kicking and screaming" into 64-bit by AMD. AMD used 64-bit as a marketing tool to regain the lead in CPU architecture (though it was temporary). Intel and Microsoft had no choice but to support 64-bit lest they be viewed as lagging behind technology.
There was very little real need for 64-bit CPU architecture, and there still isn't. Most programs do not need over 3GB of permanently mapped memory. Programs can re-map portions of their 3GB to access any amount of physical RAM. Very few calculations require 64 bits, and 32-bit CPUs can perform 64-bit calculations using multiple instructions. Even in 32-bit mode the MMX and SIMD instructions support 128-bit calculations (256-bit recently). Instruction optimization on 32-bit CPUs makes 64-bit calculations happen about as fast as they do on 64-bit CPUs.
So, pretty much any actual limitations of 32-bit software are artificial and apply to Windows rather than Linux. Some limitations of 32-bit hardware and software have occurred because 64-bit exists. They probably would not be limitations if 32-bit hardware continued to be developed. For example, there is no reason why PAE could not be expanded to allow more than 64GB of physical memory on a 32-bit CPU. If Microsoft had not produced 64-bit versions of Windows, they might have added full PAE support to 32-bit versions of Windows.