I might be mistaken, but the only seperation you can do with 32bit x86 machines gfx card ram is to, in total (system ram and GPU ram) go beyond the 4gig addressable memory. Typically 4 gig would be the limit, so system ram would shrink in favor of the GPU.
With things like pae however some systems went beyond this limit.
This of course isn't required for the vast majority of hardware these days as most things are 64bit.
In regards to csppc/a4k, its tricky because of the (frankly crap) way memory is addressed. There's only 128meg space addressable to the CPU. The space directly above this is used for things like Zorro (hence things like zorram).
Your idea, could in theory work, but I don't think it'd be much faster (possibly slower) than zorram, as while the bus is 25mhz, and the ram 64bit, the bus (correct me if I'm wrong someone) is only 32bit.
There's also the fact that the csppc firmware is hardcoded with the 128meg limit. This has been patched before trying to bypass the restrictions, but it simply doesn't work due to other hardware limitations.
Long story short, there's simply no way to get more than 128meg system ram. At least not in a way that won't cripple performance.
p.s. take this with a grain of salt. I'm going by memory here. While details may not be exact the crux is, I believe, pretty accurate.
While not ideal something like VMM or Gigamem might be your best options for the moment.