Just wanted to ask about CPU model and usage. It seems that the performance highly depends on the CPU.

Running 470x2 on old single core gave me 7 per thread on 470.
Slightly better mobo + CPU gave 14/thread (28 per 470) - almost your result, see below.
Claymore's gives 45 Sol/s per 470 on Windows - on the 1st single core machine with 1GB of RAM (45 vs 14 Sols/s of this miner on the same mobo under Linux).

The results with single card and few cards on the same machine are also quite different. I use G3240 dual core CPUs on my rigs and it is not enough to run this at full speed.

So as for me, moving some filtering from CPU/bus into cards should be considered as a priority task if we want to outperform Claymore's miner. I've read that it is ALREADY faster. But trying on my cards (7950, 7970, 390, 390X, 470) - all with G3240 CPUs - I was not able to get the same results. Was close to with a single RX470 (Elpida RAM, 1500 straps, 2000 mem/1260 GPU - 46-47 Sol/s, Ubuntu 16.04 + amdgpu-pro 16.40) but without latest eXtremal patch.

So I do not understand how do you get a half of that only (148/6=25 Sol/s) on you rigs, but almost sure that the CPU usage is the reason. Have no other ideas.