> and prove the manual wrong on the machine language level

I'll be the party pooper here, I guess. The manual is still right, and no amount of reverse-engineering will fix the architecture AMD chose for their silicon. It's absolutely possible to implement a subset of CUDA features on a raster GPU, but we've been doing that since OpenCL and CUDA is still king.

The best thing the industry can do is converge on a GPGPU compute standard that doesn't suck. But Intel, AMD and Apple are all at-odds with one another so CUDA's hedged bet on industry hostility will keep paying dividends.