Feature Request: Co-Processor, Neural Processing Unit Support.

[EndlessEden]

While the title summarizes most of the request. Supporting NPU's used in consumer and research devices, would reduce dependence on GPU workloads. Allowing lighter-weight hardware to perform the necessary calculations instead.

Such examples as the "Gyrfalcon Lightspeeur 2801S" often bundled with the ARM Rockchip RK3399 SOC. Even with a fast GPU on the PCI-E lane, not utilizing the existing NPU on-board just tosses away the potential performance gain.

Even in the worst case, it gives alternative methods for testing and exploration.