The compute machine running below the programmable shader abstraction of GPU programming models no longer looks much like it did in the 90s. It gained an integer branch unit for cheap control flow. "Shaders" can now spawn other shaders or write to the network. There can be x64 cores running in the same address space, cooperating through atomic instructions.

Said abstraction is built on open source software. The ISA is documented. The driver is in your linux kernel already. Upstream LLVM knows how to emit machine code for the chip. If you want to mangle the existing stack out of all recognition the raw material is all on github to start from. If you disregard the official vendor supplied software stack entirely the machine still computes exactly what you told it to.

This talk will sketch what the hardware looks like as an embedded target and how you might go about bending it to your will. The limits of shader programming are not the limits of our world.