We also see that when connected directly to the virtual machine through DirectConnect (bypassing the horizon server), it improves noticeably, but still not enough.
we are pretty close to the same setup .
are you on the most current driver ?.. our Esxi hosts are on 440.43 .. ie rev 10 .
We use shared direct - vGPU.. mostly M60's and P4's ( no p40's yet ) .
We use the Q profiles... i think A profile is compute mode ?.. im not sure what B mode is . ( i.e. the one your using ) .
We were about to upgrade from Horizon 7.6 to 7.11 .
Have you been on 7.11 for a while ?.. i noticed Horizon 7.12 had some vGPU 'fixes' .. however it also appears that rev has been out for about 20 mins or so ... not a fan of 'bleeding edge' .
what does nvidia-smi tell you ?... do you have ECC mode on maybe ?.. is there a compute mode for that card ?.. i think they stopped that on the M60's i was told. IE the P4/40's auto choose .
Does it sound like there are 7.11 issues w/ vGPU's maybe ?