With vSphere 4, we introduced a new paravirtualized adapter—VMXNET3. This was designed from the ground up for high performance and supports a bunch of new features.
So how does it perform? One of our engineers pitched it against the previous generation—VMXNET2. VMXNET3 not only performs better (greater throughput on transmit and receive), but consumes less CPU. These results were most significant with a 10GigE NIC on Windows (Windows Server 2008 Enterprise). The results for IPv4 and IPv6 are shown in the graph below.
Windows 10GigE throughput (IPv4)
Windows10GigE throughput (IPv6)
The results are published in this paper posted on vmware.com. It will also appear on the Network Technology site (vmware.com/go/networking) in the coming week.