They moved from HBM to dedicated silicon and only got a 48x speed up? That is so, so, so much less than I would've expected. Any numbers on how it scales?