GPU/CPU collision course -- bad?

[Intel17]

I don't know about you guys, but I'm really starting to doubt the whole "CPU/GPU" convergence idea. When we start having GPUs on CPU dies, then we lose a lot of the flexibility that discrete boards give us. There's not a snowball's chance in hell that an integrated GPU on a CPU would ever be as fast as something dedicated like a 5870 or a GTX 285.

So, are we really going to see a convergence at the high end? Am I just not understanding what's going on?

[Revs]

I think it's the other way round, as in GPU doing the CPU's job. Hence GPGPU (General Purpose GPU). It's already started happening with the emergence of tech like PhysX and Cuda, etc, made possible by unified architecture.

As for whether it'll be better or not, the market will decide. If it doesn't perform as good as current tech, people won't buy it. Therefore it's in their interest to make it faster/better. From what I can tell it could mean a leap forward in general computing performance, as GPU's are far more capable than CPU's. We just need to get through the in betweeny bit. It'll br a few years yet.

[nekrosoft13]

intel also plans to release CPUs in near future with GPU on die

[a12ctic]

intel also plans to release CPUs in near future with GPU on die

They released their first wave of them yesterday, not very impressive performance wise (Intel just slapped a 3100 on die), but fit a very tight TDP rating. It'll be interesting to see what happens once AMD introduces its fusion chips.

[shadow001]

It's hard to say,but i don't think that ultimately it's really possible to have a single processor that does it all,since X-86 chips are specialised in small,out of order execution of general purpose code,which there's thousands of applications out there.


GPU's are optimized for large,extremely parallel code,so it's a different environment altogether,and a lot of applications simply aren't suited to be processed that way,so the most i can see is having both an X-86 core and a GPU core built into the same die,which is possible with extremely good performance,as there's a huge increase in communication speeds between both cores when inside the same die,as the internal bus between both can be made as wide as it needs to be.


People are impressed with PCI-e 2.0 running at 10GB/sec,or the QPI bus,handling i think up to 25GB/sec,well that internal bus can beat that 10~100 times over quite easily in bandwith,and have a huge decrease in latency since the processors are right next to eachother,and not having to travel across a motherboard in the first place,in order to communicate with eachother.



Add that once you hit the 22nm fabrication process,overall processor designs using 5+ billion transistors are doable while remaining cheap to build,as the die sizes aren't larger than with current processors,so designers can split that transistor budget up anyway they want between the CPU and GPU cores on that die,and the end result is something extremely powerfull anyhow.


Will we be witnessing the death of the add on graphics card?....I believe so,at least for the low end and midrange products,and maybe even the high end too,since it costs a lot of money to develop those high end products,and the markets are much smaller for those expensive 500~600$ cards,than they are for the midrange and low end products,which sell a lot more in volume sales.

[Viral]

It's more for GPGPU than gaming. Unless they try and phase out expansion slots, you won't lose the option to go discreet and just have the APU(CPU/GPU) do all the usual things your CPU would.

Laptops and small form factor desktops have taken over though. It's not really that possible any more for some joe blow to buy a computer and throw in a video card when they see a game for PC they like released. For these portable and small platforms, APU's will take off as early as next year.

[rhink]

convergence won't really affect the gaming market. It's more for small/cheap- laptops, low-end desktops, netbooks, smartphones, etc. In those areas it definitely makes sense. For stuff where graphics performance matters more, there simply won't be enough bandwidth to keep the graphics chips fed. Graphics chips currently consume 70,80+ gigs/sec of bandwidth, CPU's are what, < 10 gigs/sec or so? Then you're going to fit the GPU and CPU on the same die and share that bandwidth? Riiight.

The other angle is of course GPGPU. Having a highly parallel "coprocessor" on the same die might be nice- but again, need to keep the chip fed. And we're not really there from the perspective of a GPU being a coprocessor for the CPU. Fermi might get us closer, not sure. Doing this makes a lot of sense for low power stuff, where you can get away with having a crappy CPU with a GPU picking up the slack for HD playback, that sort of thing. Higher end stuff... ehh, there really isn't a killer app yet for GPGPU, until then, it's just not going to take off.

[shadow001]

convergence won't really affect the gaming market. It's more for small/cheap- laptops, low-end desktops, netbooks, smartphones, etc. In those areas it definitely makes sense. For stuff where graphics performance matters more, there simply won't be enough bandwidth to keep the graphics chips fed. Graphics chips currently consume 70,80+ gigs/sec of bandwidth, CPU's are what, < 10 gigs/sec or so? Then you're going to fit the GPU and CPU on the same die and share that bandwidth? Riiight.

The other angle is of course GPGPU. Having a highly parallel "coprocessor" on the same die might be nice- but again, need to keep the chip fed. And we're not really there from the perspective of a GPU being a coprocessor for the CPU. Fermi might get us closer, not sure. Doing this makes a lot of sense for low power stuff, where you can get away with having a crappy CPU with a GPU picking up the slack for HD playback, that sort of thing. Higher end stuff... ehh, there really isn't a killer app yet for GPGPU, until then, it's just not going to take off.



Well,CPU's from both Intel and AMD now have a built in memory controler and in the case of the i7 platform,already use a tripple channel DDR3 memory setup on the motherboard,and rumors suggest that AMD's bulldozer goes one better and implements a quad memory channel setup,so users have to use 4 memory modules together.


That adds up to a decent aamount of memory bandwith right there,and i wouldn't be too surprised to eventually see a desktop version of GDDR 5 memory supported on future motherboard products,which the main advantage isn't the higher clocks,but the quad signaling used with that type of ram,unlike DDR/DDR2/DDR3 which uses a dual clock signaling scheme in it's operation.


I'm not saying that it's an easy thing to do overall,but there are no major stumbling blocks to overcome in terms of technology to make this happen eventually.

[noko]

A GPU/CPU design could have dedicated fast ram besides system ram for access when needed. Where for graphical operations it would be similar to like what we have now for video cards and for GPGPU operations a very fast area for processing. So performance capability could actually be higher with a GPU/CPU combo due to bandwidth within the chip itself. Problem is one could really make a mega big dedicated GPU and separate CPU that should be able to outperform the duo but at a higher cost. Computer industry has shown intergration over time always wins out so this may indeed be the future for both GPU's and CPU's.

[josiahsuarez]

with many PC games being just primitive DX9 unreal engine 3 level trash ported from xbox you could be right

[shadow001]

A GPU/CPU design could have dedicated fast ram besides system ram for access when needed. Where for graphical operations it would be similar to like what we have now for video cards and for GPGPU operations a very fast area for processing. So performance capability could actually be higher with a GPU/CPU combo due to bandwidth within the chip itself. Problem is one could really make a mega big dedicated GPU and separate CPU that should be able to outperform the duo but at a higher cost. Computer industry has shown intergration over time always wins out so this may indeed be the future for both GPU's and CPU's.


Well for a start,there are already faster communication protocols than PCI-e express 2.0....AMD developed the HTX protocol wich uses a specialised slot which accelerators can be plugged in,and this particular protocol uses hypertransport,and in 32 bit dual channel configuration,already yeilds about 42GB/sec of bandwith,bi directional of course,and it's a point to point protocol,so there's less latency than with PCI-e,which is split between multiple physical slots.


Main point is that there aren't major technological hurdles to overcome to make it work,and overall,it starts to make sense why Nvidia are emphasizing GP-GPU environments with Fermi and less so on the gaming front,since i think it's inevitable that add on graphics cards,primarily for gaming,have their days numbered.


Intel as we all know,has it's own CPU division and even though the current version of larrabee is dead,the software SDK is still being distributed,so a more advanced version will be released eventually,and AMD has both a CPU and GPU sides with the aquisition of ATI,so both companies can develop entire plaforms....CPU,motherboard chipset and GPU on their own.


It's no secret that Nvidia has been hiring a lot of former transmeta employees since 2007,specifically engineers specifically involved in CPU development,and we now have the FTC jumping on Intel's case to allow more companies to have an X-86 licence(among other isssues),which legally allows them to make X-86 CPU's,so it could turn into a 3 way platform battle.

[moozoo]

Once a decent multicore CPU is joined with a decent GPU , maybe after the next process shrink (22nm?), the need for the average person to buy a add in card disappears.
(Think AMD tri-core with a RV870 on the one chip).
Most people are getting add in graphics cards because the motherboard ones (read Intel) are so bad.

That combined with a slow loss of the PC games market to consoles (and the rise of fusion based consoles) leads me to forsee the end of the add in graphics card.
I doubt the enthusiast PC market alone will be able to sustain the development costs of seperate GPUs.

The bad thing about this is that without add in GPUs you can not scale up to insane performance.

[JasonPC]

Until they can get high end performance out of these CPU/GPU chips, I don't see an end to discrete graphics. I think they won't even approach the performance of the upper mid-end with these for quite a while. There is a market for high end outside of gaming and that's what nvidia seems to be working on so much with Fermi/Tesla.