ptxas never issues an HMMA.16816 within 8 cycles of the one before on
sm_86 (16 on sm_121: the tensor pipe's rate); the model asks every
instruction after an HMMA to wait 8, which is stricter (ptxas fills the gap
with other work) and puts a product's dependent, however few products
apart, past sm86TensorLatencyCycles.
553def sm86TensorIssueStall : Nat = 8The compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.