AMD is making Fusion chips from x86 CPUs with on-board graphics accelerators (GPUs). Current GPUs have up to 360 Gflops of processing power. The Fusion chips aim to increase performance-per-Watt for applications such as 3D graphics, digital media and technical computing. There are two significant design challenges in developing such architectures — power management and memory hierarchy.
By 2010 AMD hopes to release a Teraflop CPU/GPU processor. He showed an example of how this would be achieved: 48 GPU pipes x 8 Flops/cycle x 3GHz = 1 Teraflop (peak performance) per socket. AMD also announced that it will release a processor combining the CPU and GPU (graphics processing unit) on one die in late 2008 or early 2009. AMDs CTO Hester said that the new CPU/GPU processors would offer a good Microsoft Vista experience as a minimum. Hopefully we can come up with better uses for teraflop computers than running a bloated operating system.
Note: that the first teraflop supercomputer ASCI Red started broke the teraflop barrier Dec 1996. So it may take 12-14 years for that to migrate to personal computers.
A personal petaflop computer would be expected in 2018-2021 given the same gap as the teraflop. I have predicted that it will happen sooner in 2016-2018.
GPUs have been improving at 2 to 2.5 times every year which is faster than the 1.5 times improvement of Moore’s law The 2.5 per year rate against the current 350 Gflops/second GPU performance would mean petaflop performance by 2015.