CWSilicon Co-DesignChips and infrastructureWhy More GPU and HBM Bandwidth Doesn't Solve AI's Data Movement Bottleneck