For a lot of NN-based analog modeling you don’t even need the GPU. I trained a model on the API 550A EQ and SIMD is fast enough for real-time inference.