NV11 10de:0110 (AGP, 32MB, Dsub, TVout) softquadroable into 10de:0113
NV11 10de:0110 (AGP, 32MB, Dsub)
NV15 10de:0150 (AGP, 32MB, Dsub, TVout, TVin) softquadroable into 10de:0153
NV18 10de:0185 (AGP, 64MB, Dsub) softquadroable into 10de:018b
NV20 10de:0201 (AGP, 64MB, Dsub, TVout)
NV28 10de:0281 (AGP, 128MB, Dsub, DVI, Tvout) softquadroable into 10de:0289
NV30GL 10de:0309 (AGP, 128MB, DVI, DVI)
NV34GL 10de:032a (PCI, 64MB)
NV44 10de:0221 (AGP, 128MB, Dsub, DVI)
G86 10de:0XXX (PCIE, 256MB, Dsub, DVI)
TODO
see if the multisampling technique from geforce 2 can be used for hardware mipmap generation
Install an older driver on the nv28 (3xxx, 4xxx) to play with ext_vertex_weighting (figure out if it has a second modelview matrix for vertex weight stuff)
run test_startup() on the nv18 in nv18 & nv18sgl modes
fix the multiple fifo free issue (it seems fifo_free is called multiple times)
look at the nv15 with tv-in. see if it has a bigger aperture, which would be where the tv in regs sit.
add support for a crash hook for ubuntu automatic bug reporting.
check that my tnt2m64 has proper detection of its vram amount
explore the usage of movnt* to write to PCI memory. This will avoid polluting the cache, both for DMA source memory and FIFO writes. This might explain the performance difference of PCI and AGP.