T4 vGPU(in fixed share scheduler): vGPU's performace is different when another vGPU is free or running

i create two T4-8Q vGPU from Tesla T4, and deploy them on different virual machines.
Then i run gpu-burn to get the performaces(Number of executions per unit time) of vGPU and GPU.
And i find out the performace result is strange. For example, the result is just like below:
1: T4 passthrough mode: 13000
2: T4-8Q vGPU in fixed share scheduler:
a) vmA running, vmB free: 9000
b) vmA running, vmB running: 6000

I was expecting 2a’s result is 6000 but the actual value was 9000, which makes me confused.
can anyone explain this? thanks a lot.

6000 is in line with expectations(6500 loss 500 after virtualization).
The virual machines os is centos7.9, and vGPUs has been liscensed.
gpu-burn is an open source program which run above cuda.