10-19-2026, 02:15 AM
Interesting, I hadn't tried that. That matches what I saw on 5090. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.
measure twice, benchmark three times