Yeah, clKernel 3  its uses double memory.. now if your running gpu with 16GB Memory i assume you would be fine.. but if like me alll my cards are 8GB and the ETH DAG is a little of 4GB for clKernel 3 to unpack the dag acrosss all your rig's gpu's i guess to work in dual thread mode is how I imagine it works on a 8GB system your out of luck.  But that's not to say its impossible for it to run in clKernel 3 as I have managed to get my rig to run partial clKernel 3

i.e

clKernel 1,1,1,3,3,1,1,1,3,3

specifying for each gpu until the miner fails to initialize / generate the dag's.

As for further reading I stumbled on a similar issue whilst mining XMR. (Cryptonote/Night) before change to (RandomX) *I Think it was XMRig* i forget which miner it was but if memory serves me right there was some info on it over on their (github) miner repo..

If you like I'll dig through my collection of old miners/tools to see what miner it was and then I can look for the repo which contained the info i think you require on how miners handles the kernel / bins. Im sure it was xmrig or xmrminer.. (I forget, but can check ..)


* It could be in here (https://github.com/xmrig/xmrig-amd/blob/master/README.md#usage)
* Old Claymore's XMR Miner Here (https://github.com/nanopool/Claymore-XMR-Miner)

OR deep in back of memory maybe it was sgminer where I read something about clKernel and Kernel handling / types....

HTH's