64-bit leap?

Welp.. Got a new 6800 Ultra in the system. Beautiful card, runs faster, but I'm damn sure I got a pretty bad bottleneck with my 3200+. Why? Because although 90% of the time it's 2x faster than my old 5900 Ultra, there are still places in games (Doom3 and Half-Life2) that run at 30fps. Technically, that's not all that bad, but what bugs me is it runs at 30fps REGARDLESS of the quality settings. I don't have vsync on. It runs the same speed at 640x480 lowest settings in Doom 3 in a certain area as it does at 1024x768 4xAA 8xAF, which tells me my system is choking. The question is: would I see much of an improvement if I upgraded to a 64-bit 3200+ on a dual channel motherboard (I have 2x512 sticks)? I know the processor is no different in raw megahertz, but with a dual channel setup and and direct path to the processor instead of an FSB could I expect much improvement? I'm kinda hesitant to shell out $300+ if it's not going to make a difference. Opinions?

>>5
64-bit is "faster" for two reasons. And this assumes two things: (1) We're comparing x86 to x86-64.
(2) We're in native mode for each. IE we're comparing identical programs that are compiled and optimized for each processor.
 - You only need one 64-bit register to hold a IEEE double -- that alone will make anything using doubles much faster.  There will also be no reason not to use 64-bit words, so trivial changes could give huge speedups to algorithms that were originally designed for 32-bit processors.
 - Eight general purpose registers were added. This turns the x86 from a stack machine to something that can utilize its cycles much more efficiently.  I suppose they could hardwire the reverse compatibility mode to have the first 48 bytes (or whatever) of the cache be stored in the unused registers, but I'm guessing that would be a little too complicated.
 - "64-bit is usually slower since your memory structures are suddenly larger and take up more system bandwidth" is just ignorance. If there are 64 data lines and 6 address lines, the "bandwidth" is negotiated in 64-bit blocks anyway, which makes it just as fast as though it were 32 bits. They're still working at the same clock speeds. That's like arguing that 8-bit is faster than 32-bit for the same reason. Wow, if it's 8-bit, the "memory structures" (people who aren't idiots call it "alignment") are only one byte!

So, to answer the original question, it's not quite worth $300, but it would be faster if it were compiled specifically for x86-64. AMD says that there would be an immediate 25% performance gain thanks to the integrated memory controller, but I'd put more trust in actual benchmarks.