In instances of detected low GPU utilization, GPU Boost drops kepler clocks to stock speeds. This results in further FPS loss in the following situation:

Specifically, in Tera and Guild Wars 2: two dx9 apps that utilize the Umbra 3 Middleware. Umbra 3 has a known cpu bottleneck when there are many player meshes on screen and occlusion culling is being actively calculated. This results in lower GPU utilization.

However, performance is degraded further when GPU Boost reduces the ROP domain clockspeed of my 680 GTX to 1006mhz from my overclock of 1267mhz. MSI Afterburner graphs show a correlation between the additional fps loss and clockspeed reduction along the same time course.

I compared to my 580 GTX where I am able to force a static clock. In the same situation where utilization drops, the impact on FPS is much less severe on the fermi. (10-15 fps difference!)

I understand this problem is not directly related to nvidia and the overall performance degradation is attributable to cpu bottlenecking in umbra. However, further FPS is lost on Kepler cards in this situation when GPU Boost detects low utilization and drops the clockspeed to stock.

I understand that the design of GPU Boost is intended to manage power/utilization automatically, but this design removes a degree of control from the end user. The issue with this design is that there are other factors where a reduction in ROP domain clockspeed can cause FPS loss independent of the current GPU load.

The umbra occlusion culling cpu bottleneck is a perfect example of a situation where users will see more FPS on non-kepler cards where a static clock can be forced.