did you ever imagine owning a PC powerd by a CPU capable of 512 gigaflops to 1 Teraflop, this sure sounds exciting, but how will this be done? well it could be done in the following ways:
1) A GPU next to the CPU can greatly increase the amount of FLOPS when properly used.(Fusion)
2)A vector engine can be implemented onto the CPU. This what Intel is actually doing for Nehalem, they
are placing a vector engine inside the CPU!, somehting like th Alpha EV-9 proposal.
I don't really have enough evidence to support Intel's approach, since this is all based on rumours around the web. However, I do have enogh information about the AMD fusion project. The fusion GPU is programmable and can be used for a wide range of massively parallel applications.
Although the two approaches are somewhat different, AMD and Intel have the same goal in mind, combine vector processing and multicore to allow CPUs to achieve more flops. Since the fusion GPU is designed from ground up to support gfx, using it for HPC will result in wasted silicon since only a proportion of the GPU can be used for taks other than graphics. Intel's vector engine has hte potential to be fully utilised for both HPC and desktop apps, but since game developers are already used to GPUs, and all the software at the time of Nehalem would only be able to utilise GPUs, so a vector engine for a CPU would be somewhat useless for desktop, but once programmed for, will provide an excellent solution, just like the VMX engine for IBM PowePC.
1) A GPU next to the CPU can greatly increase the amount of FLOPS when properly used.(Fusion)
2)A vector engine can be implemented onto the CPU. This what Intel is actually doing for Nehalem, they
are placing a vector engine inside the CPU!, somehting like th Alpha EV-9 proposal.
I don't really have enough evidence to support Intel's approach, since this is all based on rumours around the web. However, I do have enogh information about the AMD fusion project. The fusion GPU is programmable and can be used for a wide range of massively parallel applications.
Although the two approaches are somewhat different, AMD and Intel have the same goal in mind, combine vector processing and multicore to allow CPUs to achieve more flops. Since the fusion GPU is designed from ground up to support gfx, using it for HPC will result in wasted silicon since only a proportion of the GPU can be used for taks other than graphics. Intel's vector engine has hte potential to be fully utilised for both HPC and desktop apps, but since game developers are already used to GPUs, and all the software at the time of Nehalem would only be able to utilise GPUs, so a vector engine for a CPU would be somewhat useless for desktop, but once programmed for, will provide an excellent solution, just like the VMX engine for IBM PowePC.
