understanding the difference in programming and high data stream parallel operation of GPU and how non parallel processes are still better handled by CPU is it possible to try to bridge the gap however by making individual cores themselves of a CPU kepler based chips in an of themselves. so each core of a cpu could be for example a 50 core kepler gpu chip. so it would technically be a core of multiple cores in a gpu configuration. obviously redesign of the gpu core configuration will be needed to handle the type of cpu instructions and calls such as virtual memory that come from say an operating system though i don’t see it as a great leap to be able to do that; as well at first glance it doesn’t seem like any large amount of new programming would be required.
what is the take on this from all of you out there?
i know a gpu core is slower however that's why it would be an actual small gpu chip to take the place of a single cpu core. is there a way to handle this with current programming languages and methods or would this possibility still be say a decade or more away because of needed adaptation of the entire operating system/application infrastructure in order to fit such a chip design? just a thought what do i know
I think you’ll just want to buy an Intel Xeon Phi.
The cheapest model has 59 mostly independent Pentium cores with a 512 bit vector ALU for you to play with.
here you have the full flexibility to let every core do what you need it to do.
the idea is more the feasibility of achieving/utilizing GPU computational speeds and bridge the gap
so to speak between the differences in processing capability of a CPU and GPU because if this can be achieved you could say have a CPU with 1536 GPU cores. using the 192 core tegra k1 as an example so each
cpu core would assuming 8 cores be the equivalent of a single k1 chip so 8 k1 chips make up 1 CPU just an example to clarify the idea
this is a thought experiment with idea of generating thought on increasing CPU speeds and capability
by taking on GPU features. in other words replacing the CPU with something faster while not requiring any vast change in current architectures and programming languages
AMD of course markets APUs (accelerated processing unit), which is a gpu, cpu mix/ blend/ hybrid…
those are used as replacements for having a gpu in laptops which misses the idea of what im trying to discuss.integrated graphics is not the subject. its speed up of cpu based processes
im not trying to talk about replacing a gpu im talking about cpu redesign to increase cpu speeds through the use of gpu computing architecture.
i am not convinced AMD would agree; they equally market the apus as HPC solutions, if i am not mistaken…
at the very principal level, AMD has done what you are proposing, if you look carefully
and even Intel’s high end deskop parts are moving into the direction of a hybrid design (CPU cores moving towards data parallel compute, and GPU cores moving towards general purpose code execution)
a) 512 bit wide AVX2 ALUs
b) more and more physical cores
c) the inclusion of graphics cores with up to 40 execution units which can be used for offloading computations through the OpenCL API.
But the Xeon Phi design is way more radical because its cores can be programmed either via OpenCL for high performance computation, or they can used as general execution units, being nearly x86 compatible (some weird exceptions apply…).