I try to make it kinda simple. For some reason, a Rebound track by certain user's of here's ( ) kinda opened me to this approach. As also Lastisland said. Saw, detune, load of reverb. I tweak harmonies at this point and then, when it needs some more harmonics I start to add some richer layers, but just very carefuly. Huge granular stuff is fun, but I feel like, especially with Omnisphere being so popular now, it gets kinda overused. Dunno.

Let's get technical. CPU works this way: It waits for a number called "instruction" which tells it what kind of operation it will perform. Immediately after it it awaits number of a same length called "operand" which is input data for that specified operation. Having both those numbers bigger means two things. CPU makers can make CPU smarter by having bigger instruction sets (a list o possible operations) and CPU can recieve longer operands. The first advantage doesn't apply to music software too much as for compatibility sake developers tend to not go crazy about new "exotic" operations CPU developers add. The length of operand is a different story, becouse memmory addresses are sent to CPU as operands. With 32-bit address you ended up with maximum of 4 gigs of addressable RAM. With 64-bits, it goes up to terra-bytes. This makes change for people using big sample libraries and multi-samples, becouse you can load much more of them into RAM before their DAW crashes. As synth-ish stuff goes, not much changes actually.

Unless specified, 32-bit and 64-bit version of software doesn't mean what bitdepth software's internal engine uses.