A GPU store takes 6 cycles — but the data's journey is just beginning
What happens when a GPU writes memory
Following one STG.E instruction through an RTX 4090, this deep dive traces what happens after a warp writes memory: the LSU, the coalescer packing 32 lanes into four 32-byte sectors, a write-through L1, and the L2 slice that acknowledges the store while the warp has already moved on. The data sits dirty in L2, and its eventual eviction to DRAM is governed by RRPV replacement and an 8-dirty-line janitor rule.
A LDG.E took 255 ns, the warp waiting all the way. STG.E takes only 6 cycles, just enough to set the store in motion. Stores have a long afterlife, as they make their way through the units, the warp happily oblivious unless it wants to read them back.