it depends on the cpu quite a bit you will find... surprisingly. unelss you go the route of using the like vertex-array-range (OpenGL), which I''ve found, if the cpu is good, has little effect on overall tris/sec rates.

as for numbers, a friend of mine had a geforce 2 mx 400 (absolute peice of rubbish in other words - about the same as my old geforce 1) on this, I was managing around the 9 million tris/sec mark with VAR using triangle strips that were optimized for vertex caching, etc (on both the mx and my gf1 - slightly less for the mx) - (about 6-7 without VAR for him, ~3 for me)

He had an athlon 1700. I have a 533 P3.

Now. today, on my same P3 + radeon 9000. With standard mixied triangles, no strips, no vertex cache optimizations (don''t underestimate these) I''ve managed to get it to pull up to 10 million tris/sec (D3D)
(this wasn''t an simple scene either, it was me rendering the nv15 quake3 map)
You should probably be able to get the same rates on a 4mx..
OpenGL, on the other hand, I havn''t written to take advantage of VAR/VAO, so it''s throughput on my system is roughly half that, but, on the gf mx, the performance of both GL/D3D is pretty much equal.
(the cpu usage is higher in GL, so things bog down quicker when lots of things start happening than in D3D.. this should change when I finally get round to VAR/VAO)