cant really tell from engineering samples,they could add more cache or higher stock frequencies ect yet
one thing would be nice to see is the temps
You can tell 99% of whats coming from an engineering sample, less so with AMD, and thats largely down to software support/compilers. Most things are done on Intel compilers which get updated long in advance of new releases while AMD struggle along with very little support. Bulldozer gained speed both in stuff recompiled for it, stuff not on windows(where software is less Intel centric(still more supported than AMD but less, avoided shall we say).
Engineering samples bar huge problems, are chips that are just early production runs, very little usually change and even the odd bug fix are rarely performance involving.
More cache enabled in certain parts, but they will never just add more cache in last second, the entire chip top to bottom is geared to a specific level of cache for optimum performance, its manufactured with that, adding more cache would mean setting back a chip a couple years.
Clock speeds are relevant, because final bins are usually clocked higher, but not always, with a dodgy process that isn't maturing well, sometimes to hit power levels a engineering sample that was at 150W that they hoped to be 125W by launch, ends up clocked down instead. But these chips were clocked at 2.8Ghz to compare, retail will be way above that, but you can overclock a Sandy to 5Ghz plus, how fast the chip is is how fast the chip is, retail clock stocks mean nowt to people on an overclocking forum.
AS others have said, Haswell architecture(actually Intel as a whole) is 99% focused on power efficiency.
You can make a chip 3 times faster by ramping up size, clock speeds, easily, and increase cost, but that chip wouldn't perform any better at 17 or 35W, making a chip with vastly better power efficiency means huge increases in usable power at those lower power levels.
Haswell should be pretty awesome for 17-35W(maybe, they seem set on shoving GT3 and expensive on die memory down everyones throats, affordable 17W chips might be significantly slower), and generally excellent for idle power and efficiency, but in terms of bringing around new levels of performance, Haswell is going to be a massive, huge disappointment.
I've said for like a year now, Steamroller is going to be interesting if we get octo chip versions as while Intel is focused on every single last ounce of power saving, AMD has done a lot for power, but a lot for performance, and an 8 core chip that is significantly faster than the current one, on a smaller process that can likely do noticeably higher clock speeds... for balls out performance Steamroller looks VERY interesting in the affordable bracket, and Kabini looks pretty awesome for power. AMD just won't get a great steamroller in a 17W bracket like Intel will with Haswell.... but Intel are going to charge through the teeth for that 17W chip so I don't really care
EDIT:- for the record, Haswell might end up looking better in real world benchmarks than the likes of Superpi, in reality the changes to micro-ops, the addition of integer only parts, is really offering versatility not increased throughput. IE when you do only integer math, Haswell won't really do anything better outside of increased accuracy in branch prediction and caching, when you have random programs doing less predictable things and asking to do fpu/int mixes that are less predictable is where the new int only parts of the core add some more flexibility and throughput... realistically this is more likely in HT situations with two programs doing different things, or more complex programs with different threads, some FPU heavy, some int. A basic benchmark like superpi would have limited benefit.