Why my compiled version seems slower than the one download?

[BVEsun]

System: Windows 10
My compiled version using cmake with:
cmake -B build -DGGML_CUDA=ON
cmake --build build --config Release

and the download one is :
llama-b3651-bin-win-cuda-cu12.2.0-x64

when I run llama-bench, I found the following:
Download One:
llama-bench.exe -m Meta-Llama-3.1-8B-Instruct-Q8_0.gguf -ngl 81
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
Device 0: NVIDIA GeForce RTX 4070 Ti, compute capability 8.9, VMM: yes

model	size	params	backend	ngl	test	t/s
llama 8B Q8_0	7.95 GiB	8.03 B	CUDA	81	pp512	5271.29 ± 38.14
llama 8B Q8_0	7.95 GiB	8.03 B	CUDA	81	tg128	52.31 ± 0.15

build: 8f1d81a (3651)

My compiled one:
llama-bench.exe -m Meta-Llama-3.1-8B-Instruct-Q8_0.gguf -ngl 81
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
Device 0: NVIDIA GeForce RTX 4070 Ti, compute capability 8.9, VMM: yes

model	size	params	backend	ngl	test	t/s
llama 8B Q8_0	7.95 GiB	8.03 B	CUDA	81	pp512	5121.40 ± 31.74
llama 8B Q8_0	7.95 GiB	8.03 B	CUDA	81	tg128	47.85 ± 0.27

build: a47667c (3650)

I have run several time and my compiled version is slower than download one.

Is there any suggestion for this problem?