Low FP16 performance for Turing w/o tensor cores (GTX 16 series)?

[pt13762104]

I have a GTX 1660 Ti, when I checked it using test-backend-ops, the FP16 performance is only 0.7 TFLOPs, 5x slower than FP32 and 3x slower than other types. When I enable GGML_CUDA_FORCE_CUBLAS, the performance of every type excluding BF16 and FP32 is 0.7 TFLOPs. This seems much lower than the rated 10 TFLOPs that my GPU have.