Just got GPU-z.
In renderperftest at 3000 objects, GPU load is at 100%, CPU load at 10%.
In that fish test, with just 20 fish, GPU load is at 70% (?!?!?!) while CPU is at 10%.
GPU usage goes up to 100% at 1000 fish, still minimal CPU usage.
So not looking too good for this being a C2 issue... Unless those developers are missing some optimizations/workarounds that Scirra is also missing.
The fact that I could render 150,000 cubes though... I mean, how can 3D be optimized so much that it's like 500 times faster than 2D?