model merge got better or just different

[griffinloom66]

I think the bottleneck is somewhere else. I get why that seems right, but on 4090 the bottleneck felt elsewhere. Could be a workload difference, but I wouldn't generalize too fast from one setup.