When searching for screenshots of 4090s GPGPU benchmark values for VRAM memory copy for comparing, there are often values with more than 20xxGB/s visible:
e.g.
In theory the 4090s VRAM has a bandwith of round about 10xx GB/s - see techpowerup and geizhals and NVIDIA (page 13) .
So the "measured" memory copy values seems to be much too high.
Because for AMD RDNA3 GPU the memory copy is round about 9xx GB/s - that fits more to the theoretical specs of the used GDDR6 of 960 GB/s:
OR:
If the AIDA64 GPGPU tool is maybe errorly measuring the large L2 Cache (72MB) bandwith of the AD102 instead of the GDDR6X RAM, then it should measure the 96MB L3 cache bandwith of AMD RDNA 3 GPU too. Maybe with a new field in GPGPU?