Of course, it works well as a tech demo and in controlled environments, and it's easy to understand and then brute-force with custom hardware. The problem is that it sucked the life out of better techniques which might actually work in the real world.
So, intel made a custom processor for this that's really good at the correlations. The original one was the D4, and later they used their movidius chip. Both have lots of multiplier-accumulate silicon, so it can do the computations in parallel. Their architectures are also set up for convolution (which re-uses a lot of data) rather than random-processing (like a CPU does), so they could feed these math engines without a lot of data transfer -- this makes it more power efficient. You could do something similar in an FPGA, but dedicated silicon is going to be faster, cheaper, and use less power.