How should we think about how much effective compute is being done with these devices compared to classical (GPU) computing? Obviously FLOPs doesn't make sense, so what does?