10-12-2026, 07:28 PM
Not sure I agree there. I get why that seems right, but on 5090 the bottleneck felt elsewhere. Not saying you're wrong, just that I wouldn't lock that in as the lesson yet.
measure twice, benchmark three times