File Allocation Tables aren't stored in memory, they're stored on the drive. Otherwise the OS would never need to reference the drive's copy of it except at boot and at shutdown (and wouldn't that be scary with Windows' reliability). Of course, an NTFS partition doesn't have a FAT; with the setting for the master file table reservation to be quite high, 35MB is reserved for the MFT on my system.
FAT32 also uses 4bytes per entry, not 2 which is what makes it 32bit. It'd be 16MB for a 137GB drive. (2^37/32768) = 4194304 clusters x 4bytes. (I'm not even sure what KF's calculation refers to.) There are also two copies of the FAT per partition. For 4k clusters, it'd be (2^37/4096) = 33554432 x 4bytes = 134MB. That's why the default is for FAT32 to use 32k clusters beyond 32GB, and why Microsoft wants you to use NTFS for larger drives.
What does the video card occupying some memory space for the on-card memory mean? The on-card memory is used for storing 3D textures and frames before they're sent to the monitor. The only time the system memory is affected is during AGP texturing. (Other than the OS having to use memory for determining what needs to be sent to the video card in the first place, of course.)
The entire nvidia driver set for version 40.72 is about 20MB. (Found by looking at the driver list in the display adapter properties and adding them all up.) I assume that much of that is loaded in memory, but some could just be referenced but not loaded until needed specifically (like when switching to 3D mode).
IRQs are only an issue based on how many controllers you have, not drives. Onboard controllers take up 1 IRQ per channel (primary and secondary). Add-in controllers including RAID on the motherboard only takes 1 IRQ per controller, no matter how many channels or drives are connected. Of course, connecting controller cards and other devices on the PCI bus is often a difficult task simply because of imcompatibilities and IRQ issues unrelated to how many devices there are.