IPU Profile

Useful efficiency = useful issue-cycle equivalents / measured tile cycles. AMP uses peak FLOPs; other kernels use arithmetic instructions or dense-copy throughput. Lane occupancy excludes known padding. Coverage is the fraction of measured cycles with estimates; cut samples, legacy profiles, and unsupported kernels are unavailable. Values above 100% indicate model mismatch. Device useful includes idle tiles over the kernel’s interval union (a lower bound with partial coverage). Sorted by total tile cycles; click a kernel to focus the profile, and click it again to clear focus.