The application performance gains resulting from migrating an infrastructure hosting multi-tier applications from a physical networking implementation to a NSX software-based one have been well documented.  Bob Goldsand and Todd Muirhead, from our performance engineering team have been preaching this gospel for some time, for example at VMworld 2017.

Implementing NSX with distributed routing and firewalling, combined with placing the virtual machines of multiple application tiers on the same hosts where possible, (and using affinity rules to keep them there), leads to the optimization of traffic flow – avoiding the physical network altogether and the hairpin flows as traffic destined for adjacent VMs transits the infrastructure unnecessarily.  The performance increase is diminished though still significant where tiers are not on the same host, and this is despite the enhanced security of micro-segmentation, providing a firewall between the vNIC and vSwitch of every VM.  The more complex, and so more tiers, the application has, the greater the gain.
While very large, 40%+, improvements were based on comparison with a traditional multi-tier network architecture, even in recently built leaf spine data centers they are significant, frequently in the high single digits.  This may not sound like much, but as each application turn (request by one tier of the app and response by another) magnifies this gain, this enables a gain of 15 to 25 percent to lead to a perceptible improvement in application performance for the end user.
Two videos have just been published, one explaining this NSX benefit in greater detail and the other how vRealize Network Insight can be used to expose these application flows in your environment: