Modern NVIDIA GPUs contain tensor cores that are optimised for MMA operations as well though right? The standard CUDA execution path isn’t comparable, but the tensor core path very much is?