You forgot their tensor core performance numbers "with sparsity".