@MartB
SFE definetely cannot outperform Cavium's proprietary offload yet. With SFE you can get around 1.3 Gbit/sec half-duplex, while Cavium is close to 1 Gbit/sec full-duplex.
I'm running LEDE on my ERL with SFE support in the kernel, and when doing SQM with CAKE you can push 200 Mbit/s full-duplex through it. Which is not bad compared to the 50 Mbit/s you get with CAKE on the stock firmware.