I've probably posted this in one of the Pro threads already, so apologies for a potential repeat, but if I recall correctly, FP32 Adds can see an ~80% gain on Dual-Issue and FP32 Add-Multiplies can see a 20% gain. With most other operations showing negligible gains, about the same perf and even penalties in some cases.
I guess the question is what proportion of these operations make up the average set of operations at a given time in a given game/engine (my knowledge doesn't go anywhere near that deep)..
But, just as a purely hypothetical eg. let's say it's ~30% FP32 Adds and ~20% FP32 Add-Multiplies; then you'd see an 80% gain on that 30% and a 20% gain on that 20% resulting in 100% + 24% + 4%. If the Flops are 18TF Single-Issue, then that gives you a rough equivalent of 1.28x18 for a more realistic estimate of ~23TF.
That said, the "+45%" rendering figure throws a spanner in the works of that, but if is referring to what we think, then that's likely a limitation of the overall system bandwidth/throughput due to other bottlenecks.