"LINPACK" (without qualification) is a double-precision benchmark, and doesn't measure raw compute flops, but what is actually achieved solving a linear system. K1 does double-precision at a rate of 1/24, and good libraries typically achieve ~90% of rpeak on similar hardware, so something around 13.7 gflops is a reasonable estimate.

https://developer.apple.com/videos/wwdc/2014/#703 around the 30 minute mark.

That's almost a "what not to do" for using Accelerate. You're introducing multiple passes over the data, and introducing extra dummy arrays that need to be passed over as well, which blows up the load-store traffic further. You're also using vvpowf to compute a simple reciprocal, which is wildly inefficient.

I don't mean to pick on you, but it's a misleading comparison. A basic transform that only gets rid of the extra working arrays and does no other optimizations is ~5-10x faster in my quick timings:

    vvexpf(&result, &negx, &localcount)
    let one = Float(1)
    vDSP_vsadd(&result, 1, &one, &result, 1, localcount)
    vDSP_svdiv(&one, &result, 1, &result, 1, localcount)