1 teraflops, 10 watts
100 GFLOPS/W? Isn't that leaps and bounds ahead of all CPUs and GPUs out there?
Apparently Quadro K6000 is capable of "5.2TFLOPs/sec for single precision (FP32)" consuming 225 W. So about 23 GFLOPS/W. Double precision for the K6000 card would be 1/3 of that (7.7 GFLOPS/W). November 2014's Green500 #1 most efficient supercomputer achieved 5.27 GFLOPS/W.
So the real story here is the power efficiency of Arista/Erista/X1, even if it's mainly for graphics performance, mobiles, and tasks like vehicular vision rather than supercomputing. If they could combine X1's power efficiency with 1/3 FP64, you could build a 33 MW exaflop supercomputer (not really though). Some cards out there also do 1/2 FP64.
Edit: According to Anandtech it is 1024 GFLOPS of FP16 operations, 512 GFLOPS of FP32. I don't see 10W anywhere, but I see what looks like 64 "single precision GFLOPS/W normalized" on page 1. Feel free to crack the code.