10-17-2026, 01:03 AM
Not sure I agree there. I get why that seems right, but on GB200 the bottleneck felt elsewhere. Could be a workload difference, but I wouldn't generalize too fast from one setup.
routing beats arguing over one perfect model