TINA: Accelerating Non-NN algorithms on AMD Ryzen NPU (2024)

Really neat proof of concept—I’m just trying to understand what tradeoffs come with bending general-purpose tasks into the NPU shape.

One thing I’m curious about: when expressing algorithms like this as static computation graphs, does that change their computational complexity in any meaningful way? For example, could certain problems end up with worse asymptotic behavior (O(n²) instead of O(n log n)) due to being flattened into a fixed graph with no branching or dynamic control?