Video Cards heating up..

[Gamingphreek]

I know GPU's are very different than CPU's. I also know that there are many many many more instructions in a GPU than in a CPU. So if heat and power consumptions ar becoming issues, why dont ATI and Nvidia drastically shorten their pipelines. Then they dont have to clock their cards as high therefore producing less heat and power consumption.
Basically take the AMD and Pentium-M route.

-Kevin

 

[Matthias99]

Originally posted by: Gamingphreek
I know GPU's are very different than CPU's. I also know that there are many many many more instructions in a GPU than in a CPU. So if heat and power consumptions ar becoming issues, why dont ATI and Nvidia drastically shorten their pipelines. Then they dont have to clock their cards as high therefore producing less heat and power consumption.
Basically take the AMD and Pentium-M route.

-Kevin

I would have to question how you "know" there are "many many many more instructions" in a GPU than a CPU. In fact, almost exactly the opposite is true -- CPUs have to be able to do hundreds of different things, whereas GPUs (with the exception of programmable shaders) do pretty much one thing -- render 3D objects. They already have heavily optimized pipelines, which are (AFAIK) FAR shorter than the ones in either the Athlon or Pentium processors.

Edit: Also, the last few generations of GPUs have been expanding the *number* of pipelines (that is, a parallel solution), not making the individual pipelines longer and more complex (well, they've added PS and VS support, but that's about it). CPU makers are now turning to a similar process by building dual-core CPUs.

 

[Gamingphreek]

Well i remember somewhere someone said that. Also you can make infrences as to howlong the instruction pipeline is. I mean:

#1 The GPU's cannot get over around 600mhz or so.

#2 The GPU's are much more complex than CPU's otherwise we would just start tacking CPU's on Video cards.

-Kevin

 

[Matthias99]

Originally posted by: Gamingphreek
Well i remember somewhere someone said that.

Well, if "someone" who was "somewhere" said it, it MUST be true.

Also you can make infrences as to howlong the instruction pipeline is. I mean:

#1 The GPU's cannot get over around 600mhz or so.

#2 The GPU's are much more complex than CPU's otherwise we would just start tacking CPU's on Video cards.

-Kevin

Keep in mind the second point I added above -- on a modern GPU, you're getting 8-16 times more IPC than the raw clock speed would suggest (less than that for vertex shader operations, although they can be done in parallel with PS operations). Plus your average modern-day GPU does more "work" (in a computational sense) than your average CPU in one clock cycle (since it is working on more specialized tasks and has dedicated hardware for most of them). It might, for instance, compute the transformed coordinate values for a triangle vertex in one GPU clock cycle -- something that, done on a CPU, would probably take dozens (if not hundreds) of micro-ops. Essentially, the GPU makers have already done what you want them to do -- they've trimmed their GPUs down to the bare minimum required to do 3D rendering, and made that one task as fast as possible. I wouldn't necessarily say they're "more complex" than modern CPUs, but they're definitely complex in different ways -- CPUs have a single very long, complicated pipeline (30+ stages), fairly large instruction sets, and multiple layers of on-die cache, as well as highly complex subsystems like ALUs, FPUs, and SIMD units to implement instruction sets like SSE. GPUs have, in essence, a large number of relatively simple fixed-function pixel pipelines (if I had to estimate, I'd say maybe 10 stages), vertex and pixel shaders, and some hardware to help with AA and AF.

If you're interested in looking into GPU architecture in more detail, AT and Tom's Hardware have both published a number of articles on it (other sites, I'm sure have done similar things -- check out Beyond3D and FiringSquad as well). It's impossible to know exact information on the number of pipeline stages, etc. (plus, you couldn't compare it directly to a CPU pipeline, since they do different things), but you can get a general sense of how the data flow works.

AT article on 6800
AT article on NV30
AT article on R420
AT article on R300

THG NV40 article
THG R420 article
THG R300 article