What's the difference between a CPU & GPU?

[latrosicarius]

I know what a CPU and GPU is, but I'm just trying to get a better understanding of the differences in the processors...

For instance, why is a graphics card better than a second CPU that is delegated by the OS to perform all graphics-related functions?

Is there a difference in the instructions like 3DNow! and stuff? Do they somehow have a better relationship with DirectX and OpenGL?

I realize that modern graphics cards use onboard GDDR3 RAM, rather than the shared system DDR RAM, but I'm more interested in the actual Processing Unit, not how much or what kind of RAM it has access to.

Thanks

[arthurleung]

CPU is general purpose (i.e. can do any instructions)
GPU is not, its only for graphics, although its getting closer to general purpose now, i.e. h264 hardware decoding, physics calculation(?)

And GPU have a lot more transistors than a CPU. But of course those GPU transistors run a lot slower than CPU transistors and geared toweard parallel processing (i.e. 24 pixel shaders)

Given its all about business I doubt AMD, Intel will step into graphics field to compete with Nvidia/ATi, and vice versa.

Its not about intergrating cpu&gpu is being less efficient or whatever, but you need one company good at both to get it integrated.

[latrosicarius]

So business practices aside, it's possible that in the future, I could have a multi-core CPU or a multi-CPU computer, and just assign one or more of the cores/CPUs to handle all graphics, while the rest do general purpose?

The benifit of this would be if I'm using desktop applications, I could have most cores assigned to be multi-purpose, but if I want to play a game, I could assign several cores to be "GPUs" for the time being.

For instance, I would go into task manager and set the "affinity" of the game process(es) to the various cores... hopefully by that time, games will be split into different threads for graphics rendering, physics, netcode, engine, etc, so I'll be able to set the graphics processes to the cores that I want to handle graphics. That would be awesome, wouldn't it?

[herulach]

pretty much no, gpus are designed to get a lot of vector operations done very very fast indeed, and a lot of them done at once too, doing it on a general processor would be tremendously inefficient.

[DanceswithUnix]

The 3dnow instructions were originally intended to do the matrix arithmetic needed for stuff like rotating a 3d model to place it in a scene. Again, a dedicated transform and light unit in the GPU has now taken over many of those uses, though I expect the program can stull use 3dnow/sse to work out which surfaces are not worth sending to the gpu to transform and render as they cannot possibly be in view (eg surfaces that are pointing away from you).

The instructions you need to run in order to shade pixels are a very different mix to the instructions required to run a "normal" computer program. As processors are very much tuned to the workloads they handle, the two will always be bad at doing the work of the other.

[DanceswithUnix]

But of course those GPU transistors run a lot slower than CPU transistors and geared toweard parallel processing (i.e. 24 pixel shaders)

Twenty transistors in a line will, overall, take twice as long to generate a result as ten transistors in a line. That would give you a pipeline stage that runs at a 2x clock speed difference, even though the individual transistors work at exactly the same speed. The lower GPU clock speed just shows that the pipeline stages have a lot of transistors in them.

[Hottentot]

The present CPUs most people use are x86 compatble ie designed to run x86 code whether its 16 32 or 64 bit version.

The GPUs are designed to run Open GL or Direct X code. This means that there are specific hardwired features in the GPU that can be called upon by Direct X/Open GL APIs. These features execute fast and although you could get a CPU to simulate (in software) these features it would be much much slower.

Saying that if you had a lot of CPU cores maybe they could run the graphics code as fast as a GPU. I suspect it would be still be cheaper to use a single GPU.

[latrosicarius]

I see, thanks for the info, guys. I was just curious as to whether there might be a better way to do things than currently, but I guess there's a reason for doing them this way.

ty

[-ChEM-]

anyone who was around when the 3dfx card came out will appreciate what a piece of **** the cpu is for handling graphics

3d cards used to be chipset based (no main processor) which was alright, then nvidia brought out the GeForce which had a proper GPU on it

so basically the whole point was to get the graphics away from the cpu in the first place and i cant see them going back

[Woodchuck2000]

Twenty transistors in a line will, overall, take twice as long to generate a result as ten transistors in a line. That would give you a pipeline stage that runs at a 2x clock speed difference, even though the individual transistors work at exactly the same speed. The lower GPU clock speed just shows that the pipeline stages have a lot of transistors in them.

Um, larger pipelines allow faster clocks on the whole - Prescott has a 32 Stage pipeline and hits 3.6GHz, The Athlon FX has about 14 stages and hits 2.6GHz. The whole point of Netburst as an architecture was to increase the number of pipeline stages to allow the whole thing to run faster.

The reasons why GPUs run at lower clock speeds are primarily to do with the design parameters and processes used - it's too complicated to go into here!

In answer to the original question, CPUs have to be general purpose processors by definition. If you were to optimise parts of a CPU to be particularly good at any one area, others would suffer. Certain areas of computing; graphics processing, sound processing and physics processing to name the main ones, can be done <much> faster by building specialised hardware. If you try to use these for anything other than their intended purpose they'll suck though.

[Scarlet Infidel]

Wouldnt comparing the normal and cpu tests in 3dmark give an idea of just how well optimised the gpu is? I think its pretty well established that while dedicating processors can waste potential processing it is overall the best solution.

For example, when not playing games im sure itd be possible to get the gpu to help with general calculations, but itd be so bad at them you might as well leave it dormant doing its general 2D duties. The same goes for the audio processor and so on.

This thread has inspired me to read up on Cell processors which i never got round to doing.

[Chadders87]

So business practices aside, it's possible that in the future, I could have a multi-core CPU or a multi-CPU computer, and just assign one or more of the cores/CPUs to handle all graphics, while the rest do general purpose?

Was this a prophecy as now we have CPU's with additional cores designated purely to process graphics; Intels HD 3000 / HD 4000 chips?

Sorry to raise a dead thread, but it highlights how the future of some technologies can change direction when you don't expect it!

[scaryjim]

Was this a prophecy as now we have CPU's with additional cores designated purely to process graphics; Intels HD 3000 / HD 4000 chips?

Well, not really - HD300/HD4000 are still GPUs, they're just stuck on the same piece of silicon. The architecture is very different the CPU cores. In fact, the opposite thing has happened really - we've taken the GPU, and started giving it more of the "normal" processes that used to be run on a CPU. And let's not give Intel all the love - AMD have done the same thing, and with much more capable graphics components

This thread was started 6 months before the 8800GTX - the first unified shader, DX10 card - was launched. DX9 graphics cards weren't great at performing general computation, because the hardware was specifically designed for shading either pixels or vertices. When DX10 launched with unified shaders, those differences were lost and you had a lot of pipelines, each of which could process floating point calculations. That made it much easier to run normal "programs" on GPUs. AMD are now moving towards an architecture and proamming paradigm where the GPU just becomes another set of computation units, accessing the same memory space as the CPU.

Here's the genuinely interesting comment in this thread (emphasis added):

... when not playing games im sure itd be possible to get the gpu to help with general calculations, but itd be so bad at them you might as well leave it dormant doing its general 2D duties. ...

That's less than 7 years ago, and GPUs were considered so bad at general calculation that it wouldn't be worth using them. Now we expect a lot of software to be accelerated by GPUs. Makes you wonder what the next big paradigm shift in computing will be...