What is the difference between GPU and a "core" ?

[Esthreel]

6990 has dual GPUs but there is 3072 Stream Processors. Do all these are in GPU? I mean compared to CPU,. it has 4 cores inside as CPU, so this one has 1500 cores inside each GPU? Or can somone put a link to map of achitecture of a GPU?
Thanks.
Also when overclocking they say core clock speed. So this is the speed of every stream Processor?

 

[7andabit]

Amd and nvidia use different shaders or they count them differently. This is not precise but 5 of Amd shaders are more or less equal to Nvidias one. But that dont make it weaker. Its just a different setup they use. Nvidia uses a complex setup against Amds complex and simple shader setup.

Google for this
isca10_ckkim.pdf

Thats debunking the cpu vs gpu myth

 

[Esthreel]

Amd and nvidia use different shaders or they count them differently. This is not precise but 5 of Amd shaders are more or less equal to Nvidias one. But that dont make it weaker. Its just a different setup they use. Nvidia uses a complex setup against Amds complex and simple shader setup.

Google for this
isca10_ckkim.pdf

Thats debunking the cpu vs gpu myth

You mean this? http://www.cs.utexas.edu/
Thanks

 

[Mr. Pedantic]

AFAIK each shader 'core' in a GPU is more like the transcode unit in Intel's sandy bridge than a proper core - it's a piece of hardware tuned aggressively to perform a small variety of tasks very well, rather than being designed to allow it to perform a greater variety of tasks.

This is not precise but 5 of Amd shaders are more or less equal to Nvidias one

Four, starting with the 6900 series.

 

[sirmonkey1985]

Amd and nvidia use different shaders or they count them differently. This is not precise but 5 of Amd shaders are more or less equal to Nvidias one. But that dont make it weaker. Its just a different setup they use. Nvidia uses a complex setup against Amds complex and simple shader setup.

Google for this
isca10_ckkim.pdf

Thats debunking the cpu vs gpu myth

Nvidia's shaders = 1 shader core is split 2 ways (high speed)
AMD pre-HD6900 series = 1 shader core split 5 ways (slower speed per shader but similar performance)
AMD HD6900 series = 1 shader core split 4 ways(slower speed compared to nvidia's shader cores but higher bandwidth per shader compared to pre-HD6900 series)


hopefully that explains it better..

as far as overclocking goes AMD uses 1:1 clocks with the gpu to shader clocks.. Nvidia's differ depending on the series you look at..

 

[7andabit]

AFAIK each shader 'core' in a GPU is more like the transcode unit in Intel's sandy bridge than a proper core - it's a piece of hardware tuned aggressively to perform a small variety of tasks very well, rather than being designed to allow it to perform a greater variety of tasks.


Four, starting with the 6900 series.

Its VliV4 thanks.

 

[canmanone]

what if they take that gpu with all those stream processors and make a motherboard and use them like you would with a cpu? would that be faster/better overall from computing standards? hope i clear on my question.

 

[Pyroja]

No. Here's a super-oversimplified explanation:

CPUs are wonderful devices that are good at doing a wide variety of things at fast speeds.
GPUs are wonderful devices that are very good at doing a very small handful of things at incredibly fast speeds.

If you have an application where all sorts of processes are going on, then a CPU is in order (Like with general computing).
If you need something that performs a few operations on a lot of data really really fast, then a GPU is where you go (this is why you're starting to see GPUs used in scientific applications).

 

[NaturalViolence]

CPU hardware has evolved in response to common application programming practices. They use techniques such as branch prediction for example to improve the performance of applications that do a lot of branching, a feature that would be useless on a gpu.

A single gpu die can be thought of as a single core or pipeline. But it is a very wide pipeline. Capable of using huge vectors with extremely high amounts of instruction level and data level parallelism. However it still only has the resources to process one piece of software at a time, thus making it technically a "single core" (except fermi gpus which can be considered multi-core because they can process multiple pieces of software simultaneously). Unlike cpus it has evolved to become highly efficient at processing shader code, which can be thought of as a huge loop looping over and over again. Most software is better adapted to a cpu architecture but some specialized applications like physics simulators may be better suited to a gpu. You should see how fast you can square root a block of floating point numbers on a gpu vs. cpu. Plus a gpu is a co-processor, it can't operate independently. A program running on the cpu needs to emit instructions to memory to tell it what to do.

 

[phide]

The operational functionality of a CPU core and a shader unit are similar in the sense that each is capable of performing one operation per clock cycle, though in some cases the issuing may occur at a frequency higher than that (meaning multiple operations can be performed by one logic unit in one clock cycle). They are both 'cores' in the sense that they're both logic units, yet as others have said, the capabilities of a CPU core and a shader unit/stream processor are quite different.