There are lots of reasons, but it basically boils down to the fact that GPUs have just become extremely good at what they do: Deep pipelines - and with everything having gone unified a few years ago, we now have deep pipelines that can dynamically load balance. Highly optimized shader units - tex->blend->fog in a single cycle, most intrinsic shader functions like dot products execute in a single cycle, etc. Efficient culling - early depth rejection tests and hierarchical depth culling == cheap and brainless overdraw reduction. CPU boundedness can come from a lot of things; driver, physics, etc. Sometimes it can make more sense to NOT offload things onto the GPU (e.g. I'm not sure skinning on Intel integrated is such a good idea), which can contribute to the problem. You'd be surprised! Most GPUs for the past 5-6 years have pretty efficient depth culling, which dramatically reduces overdraw without any effort from the app programmer. And because it's so cheap (not just in terms of performance, but also in terms of die cost), pretty much every dumb, cheap chip has it. At least, for NVIDIA and ATI/AMD... I don't have much experience with Intel's hardware, but they'd be pretty crazy to not include it.