Ivy Bridge will have only 16 EUs while Sandy Bridge has 12. That's 33% increase which isn't too promising. The micro-architecture is the same so you won't see any "hidden" gains either. The clock speeds would have to be increased dramatically to provide 100% better performance than SB.
My guess is that IB's graphics will end up being roughly 50% faster. Haswell is a new MA so Intel can do more tweaks to increase the performance.
Remember, Intel has never increased EU count more than 20-30% per generation, yet got 2x gains. EUs alone won't get you high there. Think about it, if just increasing the EUs got you 100% scaling, that would mean performance is 100% bound, and that is just a bad design.
The microarchitecture has been "same" since the GMA X3000. Not the performance though.
GMA X3000: 8 EUs
GMA 4500: 10 EUs ~1.5-1.7x faster, more caches, 2x faster hardware VS, 20% higher clock
GMA HD: 12 EUs ~1.5-2.5x faster, inclusion of a proper HSR mechanism, higher clock, better transcendental performance
HD 3000: 12 EUs ~1.5-2.5x faster, GPU Turbo mode, better transcendental performance, CPU cache sharing, CISC architecture
For Ivy Bridge we can expect not only higher clock, but possibly doubling the ROP and TMU units(every arch since GMA X3000 at 2 ROP/4 TMUs), along with a general change.
If anything, I'd bet on Ivy Bridge getting more benefits than Haswell, as the shrink affords more die size.