+1.10⁶
Specific resistance or impedance of circuits rises exponentially with the reduction of the conductor.
Smaller litho SOI assemblies are actually less specifically electric efficient than larger lithos, it's just that the die is reduced in size. It's the same as with a combustion engine: if you have a 1000cc block to work with, you can only gain so much extra HP/l by implementing the same optimizations as with a 4000cc block.
The performance enhancing measures make all the difference. Intel and nVidia use a lot of trickery to enhance performance for very specific but also limited workloads. That is a dangerous proposition in a market where the focus of performance is constantly changing, and is especially changing right now that we're on the verge of many-core computing.
Simple analogy: without special optimization, a 1000 cc 3-cyl engine can only be pumped up to a certain performance level that is not very impressive, and it will have great fuel economy. With optimizations like turbo compressors, blowers, variable timing, variable valve arrangements, multiple spark plugs, intercoolers, etc... in other words, if you add 1000 USD on performance enhancement peripherals to a 250 USD block, you can boost the specific power output of that engine from maybe 75 HP to over 200 HP, whilst still maintaining great fuel economy, but the risk of breakage and the lifespan of that engine will be greatly reduced, and you're still just buying a 1000 cc 3-cyl engine for 1250 USD. Compare that to a 2000 cc 4-cyl engine, which is 2 times as big, and although it will present less fuel economy, and the vanilla block costs more at maybe 1000 USD, it can deliver 200+ HP without expensive performance peripherals, so the price for 200 HP is 1000 USD and the difference in added fuel consumption, but you're getting more engine for 1000 USD than with the 1000cc 3-cyl engine proposition. Now if you're only doing dyno runs, and the only thing that counts is the performance at max rpm, the 1000cc 3-cyl might do better on the dyno than the 2000cc 4-cyl engine, but if you look at overall performance and flexibility, the 2000cc engine will be a better choice.
To be honest, I have not seen ANY overall performance enhancements come from die shrinks in microprocessors. The lithos are small enough to fit a huge number of transistors in a compact card with a good price. Any proposition that makes more expensive cards with smaller dies with the same number or less transistors, marketed with the argument that the dyno tests (performance per clock in synthetic benchmarks) look better or the power draw is lower (which is not necessarily true, extreme cooling measures also cost energy, the amount of extra juice drawn by the air conditioning to compensate for the much higher heat emissions of smaller litho chips is more expensive than the extra juice used by a larger litho chip with less heat emissions).
The 290x runs hot because it's made to run hot, it's actually a 2000cc engine with a blower, it loses efficiency to deliver a little better dyno run spec.