Running iperf in bidirectional mode goes slower than unidirectional

I used iperf to test throughput on box A and B, both 1Gbps NIC. I found: * A->B can reach 940 Mbps * B->A can reach 940 Mbps But with iperf -d, to test bidirectional transfers, the result is only about 670 Mbps, so what can cause this problem? Switch or router?