10-04-2026, 08:08 PM
I think the bottleneck is somewhere else. I get why that seems right, but on 5090 the bottleneck felt elsewhere. Worth testing the simpler explanation before adopting the more interesting one.
measure twice, benchmark three times