I've been hacking around with this box... some differences from the plain G.

1. Uses mips big-endian mode, so you have to make a new toolchain.  Not compatible with Linksys binaries
2. Only 4MB Flash!  I've built a kernel with squashfs and an extra mtd partition for the spare space, but don't anticipate having more than around 2MB free even if I strip the image bare.  btw. their hacked kernel 2.4.18 kernel contains bits of 2.4.19, some CVS directories, all sorts of crap, so it's difficult to strip out their changes.  They also statically link some proprietary stuff into the final kernel (GPL issues???)
3. Their kernel has lots of stuff extra like ebtables for bridging... not used by the shipped frontends but could be useful.
4. Netgear provide all the tools to create a new image, so there's no guesswork.  If you run their script on a default binary it produces an identical image to the release image, so seems to be OK.

Some info...

# cat /proc/cpuinfo
system type             : 96348GW-10
processor               : 0
cpu model               : BCM6348 V0.7
BogoMIPS                : 255.59
wait instruction        : no
microsecond timers      : yes
tlb_entries             : 32
extra interrupt vector  : no
hardware watchpoint     : no
VCED exceptions         : not available
VCEI exceptions         : not available
# cat /proc/version
Linux version 2.4.17 (root@Run-P4) (gcc version 3.1) #326 Thu Mar 3 16:49:23 CST 2005
# cat /proc/mtd
dev:    size   erasesize  name
mtd0: 00286000 00010000 "fs"
mtd1: 003e0000 00010000 "tag+fs+kernel"
mtd2: 00010000 00002000 "bootloader"
mtd3: 00010000 00010000 "nvram"
# free
              total         used         free       shared      buffers
  Mem:        14016        12140         1876            0         1332
 Swap:            0            0            0
Total:        14016        12140         1876

btw.  It uses uClinux 0.9.19.  I upped my images to 0.9.20 as that's claimed to be binary compatible..  I notice openwrt uses a much newer version - is the binary incompatibility not an issue for you?

I build an image with squashfs + jffs2, busybox 1.0, etc.  seems to work.

Unfortunately while testing I made the bootloader writeable and managed to write to it..  result: brick.  It seems the DG834GT has no ROM and if you corrupt the bootloader it can't flash a new one (the bootloader is what gives you the flashing lights if a firmware upgrade breaks - it's pretty deep brickage to destroy that).

At this point I wish JTAG cables were affordable...

Do you know what other differences there are between DG834G and GT besides endianness?

The GT is a Broadcom 6348 Mips processor, with a Broadcom switch (so its switching function doesn't use any CPU).  It uses an Atheros 108MB wireless card... looks to be a fairly standard one.  There's 16MB of RAM and 4MB of Flash - there's so much RAM because the drivers take up half of it before the OS has finished booting.  The GT looks to be quite a bit faster than the G also.

From the info on the G it's similar, just the CPU difference, the broadcom specific stuff in the kernel, and the different wireless card.  It *should* be possible to do something like I've done with the G, provided the source is available in the same way as for the GT.

The best thing to so if you're planning on doing a small conversion like mine is to enable NFS in the kernel and store everything outside the machine while you're working with it (you can even make it boot off NFS but I wasn't that brave...).  That means you can modify things without having to re-burn an image every time.