It's been a long time since I read Intel's specs on the 440BX, but from what I remember it only specifies up to 4 "rows" of unbuffered memory. The limit goes up to 8 "rows" if the memory is registered. I'm calling them "rows" because that's the terminology that appears in Award BIOS POST screens.
A double sided 256MB module with 16-18 (16Mx8) RAM ICs on it is 2 rows. These are the largest modules that can be used by the 440BX.
If you stick to Intel's guidelines, the max memory is 2x256MB unbuffered = 512MB total of unbuffered memory at 100MHz FSB.
If you use registered memory, you can go up to 4x256MB = 1GB registered.
It's so easy to break the 440BX's official limits that it becomes easy to forget what the limits are supposed to be.
My experience (mostly Asus boards, but I've had some other brands of 4-DIMM boards I might have tried it on also) has been that you can go up to a full 1GB of unbuffered memory if you stay at 100MHz. If you overclock to 133MHz, that's when pushing the unbuffered memory limit gets questionable.
Several years ago I had several Asus P2B-F and P3B-F boards. I cherrypicked 2 of them (both were P2B-Fs) and sold off the rest. Both of those boards ran long term with 4x 256MB unbuffered memory at PC133 CL2 timings with overclocked P3s at 800/133FSB and an AGP card at 89MHz. Both setups were initially stress tested at 140MHz FSB and still passed Prime95/Memtest86 like that.
Breaking a lot of rules, but their 133FSB stability was excellent, *except* for the bootup. That was the one hitch - I had trouble getting them to POST the full amount of memory.
Whenever I booted the machine it would report some random amount of memory in the range from 600MB-1GB. I'd have to hit reset once or twice until it showed the full 1GB at POST. Once it did that, I'd let it continue booting and it was perfectly stable in actual operation. It was just the bootup that was temperamental.