It wasnt so much the case of 3dfx being slow to innovate, it was Nvidia murder spree.
3dfx guys were going at steady SGI small steps pace of a 3D workstation company. Nvidia came at it with LSI Logic attitude (CEO came from AMD thru LSI) of quickly iterating and shipping silicon.
My favorite obligatory 3dfx fall/Nvidia killer mode timeline Re: Bought these (retro) hardware today
Riva 128 (April 1997) to TNT (June 15, 1998) 14 months
--Revenue 1997 Nvidia $29M, 3Dfx $44M
--Revenue 1998 Nvidia $158M, 3Dfx $202M
TNT2 (March 15, 1999) 8 month
GF256 (October 11, 1999) 7 months
--Revenue 1999 Nvidia $375M, 3Dfx $360M
GF2 (April 26, 2000) 6 months
3dfx dies here
--Revenue 2000 Nvidia $735M, 3Dfx <$200M and bankruptcy
GF3 (February 27, 2001) 9 months
--Revenue 2001 Nvidia $1.4B
GF4 (February 6, 2002) 12 months
FX (March 2003) 13 months
Lets not talk about NV1/NV2 duds 😀 After hail mary Riva 128, with tringles this time, Nvidia started working on at least two designs in parallel. One to ship as fast as possible, the other implementing good but costly ideas. At the time first one was shipping the next ones costly ideas were already cheap and so new even more ambitious project was created and so on. Nvidia also employed state of the art logic simulations - all designs were emulated on a gate level before tapeout. 3Dfx was also emulating, but on a subsystem level, Low-level emulation (LLE) vs High-level emulation (HLE) situation.
ATI was kept alive for the most of this time with B2B deals, bundling pitiful Rage chips in name brand computers.
edit: updated with more financial details