Yes, it does - at least with nVidia Turing chips.
Selecting "GPU" encoding in VRD results in a much faster completion than software only, but the CPU is still very involved.
Test hardware:
CPU - Ryzen 9 3900X 12c/24t
GPU - nVidia 1650 Super (Turing)
16 GB DDR4 3600Ghz memory
source file on 970 EVO 512GB NVME
output file to ADATA SX8200 Pro 2TB NVME
Did a test recode on a 11 min file while monitoring with HWInfo 64.
No other foreground programs were active and VRD was definitely set to use "NVEnc" encoding
The GPU engine loaded to about 27%, but all 12 cores and 24 threads on Ryzen loaded also to 60%-70%
The CPU usage was not a spike, but continued until encoding stopped.
VRD completed the test in 25 secs at 786 fps
Ran same source again, but with 2-pass SW encoding using similar specs
GPU load was 0% and CPU load was again even across cores/threads and still about 50%-70%
This time, completion took 111 secs at 177 fps
I did not have an appropriate system or time to run tests using a Turing video card and a less powerful CPU, but I'm sure it will be proportionally slower on a quad or hex core system with the same NVEnc card.
So, NVEnc GPU encoding is much faster and pretty good quality, but your CPU is still being used - a lot - and how powerful it is will have an effect on completion times also.