3 articles tagged
Published benchmark data, Apple's own specifications and one hosted price list are enough to work out when buying a Mac beats renting tokens. The answer depends on prefill, not generation, and most of the M5 numbers
PyTorch 2.13 landed FlexAttention on the Mac's Metal backend, with a documented ~12x speedup over SDPA on sparse masks. Here is what the numbers mean, how to write a mask in two lines, and where MLX still wins.
PyTorch 2.13 lands FlexAttention on Apple Silicon with Metal kernels for sparse prefill and decode, up to 12x faster than SDPA. How the API works, what the numbers show, and when block-sparse attention pays off on a Mac.
