To elaborate on the cause and impact, we have to differentiate between CPU usage and CPU utilization as they are different metrics.
CPU utilization refers to whether a PCPU is idle or not, i.e. a PCPU is utilized when anything but its idle thread is executing. At the host level, when SMT is enabled, 50% average CPU utilization can mean that all cores are being utilized by one thread or that half of the cores are utilizing both their threads. The throughput of two hosts with these two similar looking utilization examples could be very different.
CPU usage is a qualitative metric that incorporates more than just utilization, like the frequency of the underlying CPU or whether a world has to share the core with another world at the time of execution. An example how frequency can cause a difference between usage and utilization would be a 50% utilized PCPU at 60% CPU usage because of 20% Turbo Boost above the CPU's base frequency. For HyperThreading, ESXi assumes a flat 25% throughput benefit for the core when two threads utilize it at the same time.
ESXi 7.0 U3 fixed a long-existing issue in which CPU usage could be under-accounted. Before the fix, when a vCPU world migrated between different PCPUs while the other PCPU of the core was also being utilized, a rare race condition could result in some of that utilization not being charged to the vCPU world and the PCPU as usage.