I'd more interested in seeing a CUDA implementation. If they were able to get 100x baseline on the CPU, I'd expect the GPU implementation to be at least 1000x b

I'd more interested in seeing a CUDA implementation. If they were able to get 100x baseline on the CPU, I'd expect the GPU implementation to be at least 1000x baseline as most tasks are "infinitely" parallelizable (row additions, multiplication etc.)