Intel claims that it's 22nm Haswell architecture will rival today's discrete graphics cards. It's unknown if the company is talking about mid-end or high-end GPUs though, and after the Larrabee debacle we aren't expecting too much..
Haswell will utilize the same power-saving tri-gate 3D transistor technology that will first drop with Ivy Bridge in early 2012. Major changes architecturally reportedly include a totally redesigned cache, fused multiply add (FMA3) instruction support, and an on-chip vector coprocessor.
The vector process, which will work with the on-die GPU, was a major focus of the post. The company is preparing a series of commands called Advanced Vector Extensions (AVX), which will speed up vector math. It writes:
Intel AVX addresses the continued need for vector floating-point performance in mainstream scientific and engineering numerical applications, visual processing, recognition, data-mining/synthesis, gaming, physics, cryptography and other areas of applications. Intel AVX is designed to facilitate efficient implementation by wide spectrum of software architectures of varying degrees of thread parallelism, and data vector lengths.