At the annual Research@Intel Day, Intel showed off a system capable of 2 TFLOPS. The Inq says the system uses 80 floating-point mini-cores clocked at 6.26GHz.
At 6.26GHz the system uses 157W, but when you scale the frequency back to 3.13GHz the system still has enough power to deliver 1TFLOPS while the power consumption drops to 24W!
In idle, only four out of 80 cores are working, at 3.13GHz and they consume only 3.32 Watts, meaning that one FP unit eats only 0.83W at 3.13 GHz.
Now, here's the big kicker for this demo. Currently, this project is actually split in two: one project is currently integrating x86 cores into an massive 80-core monster, while another project is actually stacking of SRAM and DRAM memory on top of this Tera-Scale processing monster. When that happens, cache memory will have bandwidth measured in hundreds of gigabytes per second.