You should run the cuda 11.5 build for the best speed.

hashrate bugs in 1.28

When a solution was found,  it exited the solver immediately without scanning the whole range of nonces, but the full hashrate was added. This is caused a inflated hashrate. Worse on on low difficulty.
An index bug sometimes caused hash to be added to the wrong gpu, so if you run with a good gpu and a poor gpu in the rig, the poor gpu's hashrate could be inflated.

In 1.29 we don't send shares to the pool if the diff hash changed. On a vardiff pools (p2p style) or simulator pools this might reduce the hashrate, but we did it to reduce rejected shares on real pools like Nicehash.

in v1.36 we have increased the number of samples in the average hashrate, so you should run for a longer period of time.
We reduced the default intensity to reduce the stale shares.

Poolside the cuda_11_5 1.36 should be better, but the defaults might not.

f.ex for AMD cards
--xintensity 512 is pretty good on 2miners.com . Much better than 24.

AMD 6900xt:

xintensity 24: ~60MHASH poolside
xintensity 512: ~70MHASH poolside