DigiNews

Tech Watch by Johan Denoyer

← Back to articles

What happens when a GPU writes memory

Quality: 8/10 Relevance: 9/10

Summary

Explains the microarchitecture path for a GPU store from warp through LSU, L1, L2 to DRAM, including coalescing, eviction, and write-back. Highlights timing comparisons (STG.E vs LDG.E), cache policies (RRPV, 8-dirty rule), and the observation that kernels often copy results back to host rather than reaching DRAM.

🚀 Service construit par Johan Denoyer