His entire point boils down to "GPU Compute (of all forms) is overweighted in comparisons between the Mac Pro and other workstations".
He got antsy because from his perspective, your question definitely seems kinda out of the air. In his original comment, he's basically implying that any GPU based compute solution (so these can be plain gaming optimized GPUs, the "professional" GPUs, or compute optimized GPUs) aren't worth is for the majority of use cases.
The "second card" he talked about was the Xeon Phi, and he makes the differentiation between it and other "GPU based" compute solutions since Xeon Phi consists of a "large" (sub 100) number of relatively simple, but full blown CPU cores (for example, the current Xeon Phi is based on an old Pentium core, the next gen is supposed to be basically an Atom core). This should, in theory, make it easier to exploit parallelism.
