This slide effectively describes what makes a DSL a DSL rather than just some "fast" programming language. The fact that they are purpose built to handle specific important use cases extremely well is precisely what makes them compelling, and the fact that they cannot handle all problems that well is why they can't be used for everything. It also seems like this design concept is why we're seeing a proliferation of DSLs. As we push to optimize software and squeeze the most out of slowing hardware growth all problems need to be handled with top tier performance, but this means that more and more fragmented DSLs will need to exist to meet the demands.

This slide explains the ease of use of spatial compared to FPGA C programming (OpenCL). Spatial takes away programmers' burden of allocating buffers, rewriting loop tiling and parallel parts when experimenting different parameters. This improvement is analogous to how Halide automates expressing graphs in C.