Broadwell HD 6000 significantly more efficient than Haswell HD 4600

I'm seeing a solid per-EU performance improvement on a pure integer math kernel.  

I'm comparing a 950 MHz 48 EU HD 6000 vs. a 1200 MHz 20 EU HD 4600.  Both have similar DDR3 bandwidth.

Adjusted for clock speed, the kernel shows a 60% boost in throughput per EU.  Without adjusting for clock speed, the HD 6000 still shows +26% over the HD 4600.

The improved integer throughput is a nice feature!

The current kernel might be a little light on compute and bandwidth limited so the performance improvement might eventually be even higher as the kernel takes shape.  I'll know more later. :)