dual core GPUs...
Moderators: NeilBlanchard, Ralf Hutter, sthayashi, Lawrence Lee
-
- Patron of SPCR
- Posts: 376
- Joined: Thu Jun 29, 2006 8:37 pm
- Location: Mississauga, ON
- Contact:
dual core GPUs...
with all this talk about ATi and nV's new dual GPU products, i've always wondered...
why don't we have dual core GPUs yet? dual core CPUs have been around so long... yet i've heard nothing about dual core GPUs... putting two separate GPUs on the same PCB has been done before, and is now being done again... why haven't they been able to put two or more cores on a single die?
why don't we have dual core GPUs yet? dual core CPUs have been around so long... yet i've heard nothing about dual core GPUs... putting two separate GPUs on the same PCB has been done before, and is now being done again... why haven't they been able to put two or more cores on a single die?
Video cards do highly parallel work, you can say that the 8800GTS is "128 core", even the Voodoo 1 was multicore in a way.
Multicore CPUs, on the other hand, require explicit optimizations, because much of CPU code can't be easily parallelized.
As I understand, multicore CPUs are there because now it's easier to stick 2 parallel units on 1 die rather than design a higher frequency single-core chip. GPUs have been using this approach for years, I'm writing this on a Radeon 9600XT which is 500MHz, the 8800GT is just 600MHz but it's many times faster and needs 6 times more memory bandwidth to feed it.
And by the way, multi-chip GPUs like the 3870x2, 7950GX2 are stupid, they are just SLI/Crossfire on a single card. They require specific driver support, waste half the installed memory, and in the common rendering mode (AFR) they have fake FPS because 2 frames are being rendered at once, so it can't react to your mouse movements as fast as a single card with the same benchmarked FPS.
Multicore CPUs, on the other hand, require explicit optimizations, because much of CPU code can't be easily parallelized.
As I understand, multicore CPUs are there because now it's easier to stick 2 parallel units on 1 die rather than design a higher frequency single-core chip. GPUs have been using this approach for years, I'm writing this on a Radeon 9600XT which is 500MHz, the 8800GT is just 600MHz but it's many times faster and needs 6 times more memory bandwidth to feed it.
And by the way, multi-chip GPUs like the 3870x2, 7950GX2 are stupid, they are just SLI/Crossfire on a single card. They require specific driver support, waste half the installed memory, and in the common rendering mode (AFR) they have fake FPS because 2 frames are being rendered at once, so it can't react to your mouse movements as fast as a single card with the same benchmarked FPS.
Video cards seem to be one of the areas where bragging rights takes over from actual usage. I like the video cards on HardOCP because instead of doing synthetic benchmarks where they pick one resolution and mode then report the framerates for various cards, they instead actually play the game and report the highest resolution and quality settings that you can use and still have an enjoyable gaming experience. That is a far more honest way to do an evaluation, IMHO, and ranking actual playability means that if you're getting falsely reported framerates to a benchmarking tool, it's irrelevant because they are doing an evaluation of true gameplay.
AMD/ATI have been working towards the R700 for some time, simply by selling and implementing 2 card Crossfire they have worked on the hardware and software for some time.
The R700 design was from the ground up designed to be a crossfire part, specifically 1-many cores on a single PCB, and of course a very simple step down the road would be multi-core die's.
The R700 is going to be the biggest change in the way Graphics cards work for 10-years. ATI's theory was that they could produce a very small die by the million and use 1, 2, 4, 8, 16 etc on a single PCB, much like the AMD Opteron was designed to do the chips speak to each other directly and they all have their own memory interface and are designed to scale very well.
This design is vastly different from all previous designs (except the original Voodoo which is basically the same concept) where a single GPU has become highly parallel, very complex and problematic to design and manufacture. The generation after the R700, the R750 (R800 or whatever they will call it) will either be another performance jump with an improved design or it will be a multi-core GPU design, so that either less die's per PCB are needed for the same performance or more likely more performance with the same amount of die's. This is essentially an easy way to remove the current limitation of the technology, but it will probably progress in this fashion until the current design/manufacture issues come into play all over again but with many die's per PCB instead of the current 1 or 2.
Andy
The R700 design was from the ground up designed to be a crossfire part, specifically 1-many cores on a single PCB, and of course a very simple step down the road would be multi-core die's.
The R700 is going to be the biggest change in the way Graphics cards work for 10-years. ATI's theory was that they could produce a very small die by the million and use 1, 2, 4, 8, 16 etc on a single PCB, much like the AMD Opteron was designed to do the chips speak to each other directly and they all have their own memory interface and are designed to scale very well.
This design is vastly different from all previous designs (except the original Voodoo which is basically the same concept) where a single GPU has become highly parallel, very complex and problematic to design and manufacture. The generation after the R700, the R750 (R800 or whatever they will call it) will either be another performance jump with an improved design or it will be a multi-core GPU design, so that either less die's per PCB are needed for the same performance or more likely more performance with the same amount of die's. This is essentially an easy way to remove the current limitation of the technology, but it will probably progress in this fashion until the current design/manufacture issues come into play all over again but with many die's per PCB instead of the current 1 or 2.
Andy