Absolutely false. The HW vertex program support can easily outperform a CPU, especially when you take into account things like the free negation and arbitrary swizzling, and some instructions that CPUs are very bad at (LOG, EXP, and LIT).
I can’t access digit-life (the site is very slow right now), but I can access IXBT. I cannot read the text, but benchmarks are a universal language.
Now, first of all, they are using prerelease drivers, so all bets are off.
It’s DirectX, not OpenGL. I don’t know as much about DirectX, and DirectX makes it easy to write broken programs that get bad performance. MS often takes really simple concepts and exposes them in strange and confusing ways…
I’m not totally clear on what “software” means in this context. Whose software implementation of vertex programs?
What exact vertex program is being run? There are certain vertex program instructions that can be emulated fairly easily on a CPU, but there are some that are extremely difficult to emulate efficiently. MUL and ADD with no swizzles are easy to put in SSE, but try a swizzled DP4 or a LIT instruction and you will be in a world of hurt.
My basic conclusion about the article is: “not enough information.”