Pretraining of Ripple-65B Preview achieved roughly 40% model FLOPs utilization on NVIDIA hardware without custom kernels or quantization during training