Title appears incorrect; from what I can see there's no claim of this work being faster than cuBLAS in the article. There are some claimed speedups relative to clBLAS [note: "cl" not "cu"], and some references to other work which claims speedups over cuBLAS.
Agreed - the title should be changed. Perhaps the submitter saw the Nvidia cards in there and read this as an OpenCL vs CUDA kind of thing, not seeing the principal cool-factor of OpenCL : that OpenCL interoperates across all these devices (GPUs and CPUs - despite Nvidia making it difficult to find in their SDK)
Where does cuBLAS get that DP performance from? It's disproportionate.