The relative graphics claims are particularly dubious. M1 Ultra has 10% less memory throughput than a 3090. I somehow doubt that nvidia has left performance on the table that Apple is picking up, even with the power savings of being on a smaller node. On the balance of things this just seems wrong. The intuitive range of relative performance for a 3090 in flat-out single precision vector workloads should top out at 90% but more likely 75% or less.