Does anyone know at what level SSE3 improves performance over SSE2?
It's about 2 years that CPUs support SSE3 but standard SW support remains on SSE2.
What about SSE4?

This is what i get from Intel :

Streaming SIMD Extensions 3 (SSE3) Instructions.
With the introduction of SSE2, the Intel NetBurst microarchitecture extended the SIMD capabilities that Intel MMX technology and SSE technology delivered by adding 144 instructions. The next generation 90 nm process-based Pentium 4 processor introduces the Streaming SIMD Extensions 3 (SSE3), which includes 13 additional SIMD instructions over SSE2. The 13 new instructions in SSE3 are primarily designed to improve thread synchronization and specific application areas such as media and gaming. (source: Intel)

I supposed that you already read this one. And i have the same thinking as you.

It's quite hard to find information about the performance improvement in SSE3. It's just a bunch of new instructions. As they said it's to improve specific application such as media and gaming.

In my opinion, those new added instructions are only for programmers. But in this moment, programmers doesn't really need those instructions to be applied on their program, or maybe those programmers think, that the implementation of SSE3 are not yet wide spreading enough for end user. For example, most of people still using pentium 4 or AMD with SSE2.
2 years are still short for a new technology to be known by peoples. But soon, you'll see something change and/or improve in application or games appearance because of SSE3 or maybe SSE4.

Lets say Windows Vista with its DirectX 10. Maybe this OS are using the instruction of SSE3...