I pointed neural-cost at my own AlphaZero chess net and got a confident 69% efficiency that was wrong three ways. Here is what it now reports instead, and the four bugs that only showed up on a real Apple GPU.
The neural-cost v0.2.0 GPU benchmark moves from CPU roofline analysis to Apple Silicon's MPS GPU backend — measuring PyTorch and JAX across five architectures, four batch sizes, and both eager and compiled execution modes.
An empirical investigation comparing PyTorch 2.14, JAX 0.11, and TensorFlow across five model architectures and four batch sizes, measuring compiler speedups, kernel overheads, and hardware limits on Apple Silicon.
A framework-neutral Python library that estimates a neural network's FLOPs and tensor traffic, measures its runtime, and uses a roofline model to tell you exactly where performance is being left on the table.