View Full Version : Dual Core GPU's?
With CPU's all going dual core and now quad, when do you think we will start to see dual or quad core GPU's? Just think, a quad core GPU would render SLI useless. Comments?
If I'm not mistaken, that's basically how GPUs work now with multiple paths. SLI just increases the horsepower so to speak.
Redeemed
07-20-07, 04:05 PM
If I'm not mistaken, that's basically how GPUs work now with multiple paths. SLI just increases the horsepower so to speak.
You are sort of correct. GPUs aren't necessarily multiple cores in the sense of that today's CPUs are, but they are comprised of a collection of individual units- hence why GPUs are extremely parallel and very well at doing many things at once.
You are sort of correct. GPUs aren't necessarily multiple cores in the sense of that today's CPUs are, but they are comprised of a collection of individual units- hence why GPUs are extremely parallel and very well at doing many things at once.
That's what I was thinking.
Madpistol
07-20-07, 05:59 PM
So basically, if nvidia or AMD started manufacturing dual core GPU's, it would defeat the purpose of SLI, since SLI mirrors the memory across both cards. If you don't need 2 sets of memory, why do you need SLI?
Perhaps this will make Quad-SLI on 2 cards a reality... it could happen. ;)
trinibwoy
07-21-07, 04:29 PM
Multi-core GPU's and SLI don't really have anything to do with each other. SLI was always be there to give you the ability to use multiple cards.
What we have today are single-die "multi-core" GPUs. The concept of a multi-core, multi-die GPU (like Intel's Kentsfield Quad-core) is a bit strange to me. What happens to the memory controller and thread scheduling across the two dies? Load balancing is still necessary so that's additional overhead.
The only benefit I see is lower manufacturing and design costs. As far as I understand there is no performance, scalability or compatibility benefit to the consumer. Notice that CPU's are moving towards single-die (AMD is already there, Intel will get there with Nehalem) so I see no reason for GPU's to move backwards.
radekhulan
07-21-07, 06:08 PM
The only benefit I see is lower manufacturing and design costs.
Plus marketing. Some people are easily fooled by "more GPUs is better". While multiple cores do have sense for CPUs, they are ridiculous for GPUs. GPU is already massively parallel (128 "universal" units), thus adding second core makes no sense.
Redeemed
07-21-07, 06:15 PM
Plus marketing. Some people are easily fooled by "more GPUs is better". While multiple cores do have sense for CPUs, they are ridiculous for GPUs. GPU is already massively parallel (128 "universal" units), thus adding second core makes no sense.
As frequently as this has been repeated over teh intarweb- most seem to not grasp this concept. :rolleyes:
Oh teh wellz.
no please, no, i cant take another dual gpu thread..please for the love of god somebody make a sticky so we can finaly be done with this rumor...
no please, no, i cant take another dual gpu thread..please for the love of god somebody make a sticky so we can finaly be done with this rumor...
:lol2:
multi-core cpu's exist because it became too difficult to continue accelerating single-threaded execution by throwing more transistors at the problem. So they started copying and pasting the cpu's several times on the same die.
GPU workloads are quite a bit different. They're already massively "threaded", but in a little bit different way (at a much finer level of granularity, GPU threads are much more lightweight than CPU threads). GPU's can already process huge numbers of threads and massively parallel workloads without going multicore, in contrast to CPU's.
Thus, multicore (at least in the sense of CPU multicore) makes absofreakinglutely no sense in GPU's, because the reasons that drove us to multicore in CPU's don't exist in GPU's.
zeroKnots
07-27-07, 06:08 AM
Rumor? So then why does my brother's dual-6600 kick 6800 butt? (has done since 2005).
From a programmatic standpoint, we don't get to talk to different stages of these RISC processors directly. The closest we get is vertex processor and fragment processor, so all this internal 'threading' is as magic as you make it sound. But we CAN task either GPU independently in lots of configurations for dual monitor, stereo, or whatever.
If nvidia and ATI didn't black-box all their internals and we got to program down to the bare wires, THEN it would make less sense to have dual GPU's, but still a reason for them. Almost every card has dual outputs for one thing.
Two discreet pipelines with autonomous ram for each? Discreet channels to system ram? No necessity of SLI board? What's not to like?
Off topic prediction:
AMD now owns ATI.
So Intel is now making their own GPU chips, and, being NObody's bitch, Intel will release specs on their GPUs as wonderful as their CPU specs, which M$ will hate them for, but game-programmers will pledge allegiance to them every morning.
Now independant hacks will prove that just like in the 80's, combine open hardware specs with twisted minds and you get ungodly speed and innovations.
Out pops the high-speed, free, lightweight OS that drives a wooden stake through the black heart of M$ forever, and everything goes open source from then on.
:)
With liberty and justice for all.
Rumor? So then why does my brother's dual-6600 kick 6800 butt? (has done since 2005).
From a programmatic standpoint, we don't get to talk to different stages of these RISC processors directly. The closest we get is vertex processor and fragment processor, so all this internal 'threading' is as magic as you make it sound. But we CAN task either GPU independently in lots of configurations for dual monitor, stereo, or whatever.
If nvidia and ATI didn't black-box all their internals and we got to program down to the bare wires, THEN it would make less sense to have dual GPU's, but still a reason for them. Almost every card has dual outputs for one thing.
Two discreet pipelines with autonomous ram for each? Discreet channels to system ram? No necessity of SLI board? What's not to like?
Off topic prediction:
AMD now owns ATI.
So Intel is now making their own GPU chips, and, being NObody's bitch, Intel will release specs on their GPUs as wonderful as their CPU specs, which M$ will hate them for, but game-programmers will pledge allegiance to them every morning.
Now independant hacks will prove that just like in the 80's, combine open hardware specs with twisted minds and you get ungodly speed and innovations.
Out pops the high-speed, free, lightweight OS that drives a wooden stake through the black heart of M$ forever, and everything goes open source from then on.
:)
With liberty and justice for all.
Dual GPU (two physical GPU) on a PCB board is different from a dual core (2 cores on one die) GPU.
zeroKnots
07-27-07, 08:16 AM
Doh. I thought you'd call me on the loose usage of the term RISC or dual-pipelines. :)
But howso different functionally? Gigabyte uses the term 'dual-core' to describe this board even though there's twin chips.
Doh. I thought you'd call me on the loose usage of the term RISC or dual-pipelines. :)
But howso different functionally? Gigabyte uses the term 'dual-core' to describe this board even though there's twin chips.
Not really apples to apples is it mate?
Doh. I thought you'd call me on the loose usage of the term RISC or dual-pipelines. :)
But howso different functionally? Gigabyte uses the term 'dual-core' to describe this board even though there's twin chips.
Dual GPU is prolly a better description, they prolly thought it would work better since dual core was already a buzzword from the CPU side.
You dont call a mobo with 2 CPUs a dual core for example.
Dual GPU, and Dual core GPU.
Frankly I havent really seen the point.
It would prolly cost alot less transistors to expand on shader units, rops etc then to merge 2 fully fledged cores.
Maybe there are benefits, dunno really.
But the idea has been coming up then and then since Dual Core CPUs started showing up, and often by people not really understanding how parallell GPUs are in contrast to CPUs.
zeroKnots
07-27-07, 11:22 AM
I have to agree, that must be a term that someone in marketing used and it was just dirt-cheap to put two GPU's on a PCB rather than make a new die.
I'd still like to see two parallel GPU's on one die only because of multiple monitors, stereo-3D making a huge comeback, and then optionally running concurrently when youre not doing that stuff. All up to the programmer how to balance all that.
But then I'd rather have lower-level access to everything.. won't hold my breath on that except maybe stuff like this guy:
http://wiki.duskglow.com/tiki-index.php?page=OGPN17&PHPSESSID=6f20d501db8bf28bd30dfbfdbd2f0813
multi-core cpu's exist because it became too difficult to continue accelerating single-threaded execution by throwing more transistors at the problem. So they started copying and pasting the cpu's several times on the same die.
GPU workloads are quite a bit different. They're already massively "threaded", but in a little bit different way (at a much finer level of granularity, GPU threads are much more lightweight than CPU threads). GPU's can already process huge numbers of threads and massively parallel workloads without going multicore, in contrast to CPU's.
Thus, multicore (at least in the sense of CPU multicore) makes absofreakinglutely no sense in GPU's, because the reasons that drove us to multicore in CPU's don't exist in GPU's.
Well, when I say Dual-Core GPU I'm thinking more in terms of a single PCI slot card that instead of going the 7950 route and having two GPU's on it, it would have the two cores on one die or even four cores on the one die. Then toss in a physics core to boot! I see no reason that will not eventually happen.
That's exactly what I'm saying. Multiple independent GPU cores on one die make absolutely no sense for the reasons I already posted. It's possible to argue that GPU's are already 'multicore' because they possess lots of individual processing units, but it's still at a much finer grained level than cpu's.... it's a completely different paradigm, so I don't think the term "multicore" really applies.
Sticking a pair of G80's on a single chip, doing on-chip SLI, or the like is completely nonsensical.
zeroKnots
07-28-07, 04:04 PM
Theres the real issue. What programming paradigm is all it comes down to asside from better switching times on one die.
It's just simpler for me to think in terms of two identical units and probably simpler to arbitrate various loads that way.
Way off in OT deepspace.. I'm building a board with 16 atmega88 microcontrollers in a matrix. Just a hobby, nothing serious. Any MCU can talk to any other, or you can pipeline, parallel them, raster through them etc.
Or to expand further, ea 4x4 matrix can be treated like a single unit in a larger matrix.
Fairly relivant model for this discussion I'd think.
vBulletin® v3.7.1, Copyright ©2000-2012, Jelsoft Enterprises Ltd.