Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

AMD Fusion... Past or the Future...

joonkp1976joonkp1976 seoulPosts: 93Member
I am 76 born...  So I know the stinky past of the computer industry when CPU handled Graphics processing as well.  Now now, kiddies, do not go kill your parents at the truth.  HAHA...  But I always wondered how AMD could call a slogan either 'The Future is Fusion' or 'Fusion is the Future' one of either I forgot the exact.  The good old past where motherboards came with graphics chipset and like 512KB or 1MB of graphics memory.  Of that era my unforgettable being games like Air Warrior, Ultima Online, etc...  So kiddies go kill your parents for borning you inferior, later then us and happy gaming, folks...  (DON'T kill your parents though :>:>:>)


  • QuizzicalQuizzical Posts: 17,354Member Epic
    AMD marketing has abandoned the Fusion brand name.  The problem is that it ended up getting used in the Fusion System Architecture Intermediate Language, or FSAIL for short.  So you can see why one of those letters had to be changed, and the S doesn't count.
  • joonkp1976joonkp1976 seoulPosts: 93Member
    Originally posted by Quizzical
    AMD marketing has abandoned the Fusion brand name.  The problem is that it ended up getting used in the Fusion System Architecture Intermediate Language, or FSAIL for short.  So you can see why one of those letters had to be changed, and the S doesn't count.

    Thank you again for being a friend, Guide member~!  YAY, and happy gaming, folks~!

  • treysmoothtreysmooth Martinsville, INPosts: 648Member Uncommon
    Actually I attended a intel conference and they were talking of the future of cpu/gpu fusion as well. Onboard graphics and gpu's embedded into the cpu are two very different things. Gamers will always be buying external graphics cards but what if your cpu had another processor just for graphics as well? That is in essence what amd did with some of its mid tier graphics cards and some of their cpu/gpu hybrids. The interal gpu and the external can work as a cross fired set of gpu's. You get alot of bang for your buck if that's what your going for.

    The intel side is full integration of the gpu chip into the cpu. I do know that with intel you can often assign gpu's different task so it might help with that. But don't discount gpu's that are inside the cpu, with the shrinking die sizes more and more can be put into that cpu.

  • CleffyCleffy San Diego, CAPosts: 5,400Member Rare

    The reason graphics went Discrete in the first place is because you can get better performance on video rendering by having purpose built silicon for the task, and dedicated memory with quicker access rate.  Now we can take that silicon and put it back onto the CPU without heat issues.

    Right now the memory becomes a problem with the APU.  Making a custom APU on a SOC design where the graphics chip has access to its own memory could make the design substantially better as it would allow access to faster memory.

  • ShakyMoShakyMo BradfordPosts: 7,207Member Common
    Yeah you've got your ddr5 directly next to your gpu in a video card.

    Where as with an apu your using you're ddr3 sysyems RAM and having to traverse the motherboard to it.
  • QuizzicalQuizzical Posts: 17,354Member Epic

    What drives integration like that is the successive die shrinks of Moore's Law.  Approximately every two years, you do a full node die shrink, and thus get twice as many transistors in the same die area (and hence cost of production) as before.  There are a lot of things glossed over there (half nodes, different process nodes of the same size, some product lines skipping nodes, some coming faster or slower than "two years", etc.), but the general trend is that every two years, you get twice as many transistors to work with as before.

    The question is what you're going to do with those extra transistors.  Many years ago, getting twice as many transistors to make a processor meant you could take that single core processor and make it run a lot faster.  But sometime around 2004, physics got in the way and that stopped scaling very well.  You could, if you really wanted to, make a single x86 core that took a billion transistors.  But it would be really inefficient.

    So your processor has a bunch of transistors available that a single processor core can't put to good use.  What are you going to do with them?  Well, you could add a second core.  And then a third and fourth and so forth.  But doubling the number of cores every two years will very quickly get you more cores than most people have any plausible use for.

    Another route is to create new fixed function blocks in the processor to greatly accelerate certain tasks.  Take, for example, the AES-NI instructions that Intel introduced a few years ago.  They're completely worthless for everything except AES encryption, but they speed up AES encryption by about an order of magnitude.  You can likewise create new instructions for any other special purposes that you can find.  But you're not going to double your number of transistors every two years that way, as the AES-NI hardware doesn't take much die space.

    You could take functions that were previously done by separate chips and integrate them into the processor.  Rather than having two separate chips, having them in a single chip means no need for external hardware to allow the two chips to communicate.  It can save considerably on production costs, too.  AMD integrated the memory controller into their processors in 2003, and Intel followed suit in 2008.  Intel integrated the PCI Express controller into their processors in 2009.  Putting graphics chips in the same die as the processor just continues this trend.

    But if the advantages are so great, you could ask why everyone didn't just do this a long time ago.  The answer is that there are disadvantages, too, and those sometimes outweigh the advantages.

    One problem is heat.  It's pretty easy to cool a chip that puts out 50 W.  Dissipating 100 W isn't so hard, either, and 200 W is still practical on air cooling.  But 1000 W from a single chip?  Possible, yes, but customers aren't going to like the price tag on the massive phase change cooling apparatus.  Split that 1000 W into twenty chips with 50 W each and it's easy to cool again.  Indeed, racks of servers do roughly this, and can have a single rack that dissipates thousands of watts.

    Another problem is yields.  Making a 10 mm^2 die without defects is pretty easy to do.  Making a 100 mm^2 die without defects is still practical.  Try to make a 1000 mm^2 die and it's highly probable that it will be defective.  Some chips can handle large die sizes by having a lot of redundancy, rather than needing flawless chips.  Video cards do this a lot:  for example, if a Tahiti die is fully functional, AMD will sell it as a Radeon HD 7970, and if one or more SIMD engines are defective, they'll disable four of them and sell it as a Radeon HD 7950.  Having to throw 2/3 of your large die chips in the garbage is very expensive, but if most of them can be cut down a bit and sold as a lesser product, that lessens the blow.

    But that's not always practical.  Whether an AMD Tahiti chip can be salvaged depends in part on what is defective.  There's probably only one video decode block, so you can't just disable one and use another.  Try to sell that to the general public and you've got a video card that can't play back videos.  You can try to get around that by making multiple copies of everything, but that bloats the die and makes it far more expensive to produce.

    If you're integrating a bunch of other functions into a processor, then they have to all work unless it's acceptable to disable them.  If your memory controller is built into the processor is defective, then you now have a processor that can't access system memory.  That's worthless.  When integrating a bunch of things into a processor meant you'd end up with a 1000 mm^2 die size, that was impractical.  When you can integrate a bunch of things and end up with a 100 mm^2 die size, you do it.

    Another problem is IP.  You can have an Intel processor and an Nvidia video card in the same computer.  You can have an Intel processor and Intel graphics in a single chip (Sandy Bridge and Ivy Bridge).  You can have an Nvidia processor and Nvidia graphics in a single chip (Tegra).  But you cannot have an Intel processor and Nvidia graphics in a single chip unless Intel and Nvidia agree on the terms of building it.  Licensing IP does happen sometimes; ARM's entire business model is that they'll design processor cores and graphics and let basically anyone who wants to integrate them into whatever chips they want, for a suitable fee.  But there are still obstacles to overcome even when both parties are inclined to make it happen.

    Yet another problem is process nodes.  There are different process nodes designed for different purposes, optimized for different clock speeds and levels of power consumption.  TSMC alone offers something like six different 40 nm process nodes.  If you want one chip that clocks extremely high (say, an x86 desktop processor) and another that clocks very low while needing extremely low power consumption, you can get a process node geared for either one.  But there isn't a process node that is good for both, which is what you'll need if you want them in a single chip.

    Another problem is I/O, and this is a physical size limitation.  If a chip is a given size, that limits how much size you have for output from the chip.  This is becoming an increasingly severe problem as successive die shrinks double the number of transistors without comparably I/O capabilities.  A DDR3 memory channel needs 240 pins.  Socket LGA 1155 has 1155 pins.  Socket AM3+ has 942 pins.  If you've got an awesome chip for one of those sockets but it needs eight DDR3 memory channels, then you see the problem.  They also both need a lot of pins for purposes other than memory (communicating with the chipset, power intake, etc.).

    In some cases, integrating more functions into a single chip can actually help with I/O, if it means two chips that would have needed a ton of pins to communicate with each other now don't need any on either end.  But that doesn't always work; you could skip the memory pins entirely if you built the system memory into the same chip as the CPU and GPU, but SDRAM is built on a very different process node, and that would bloat the die size to something enormous and impractical.  Well, it's completely practical if you only need 8 MB of memory or so, but not if you want 8 GB.

    One final problem is customizability.  Suppose that ten people each want to buy a computer, and it needs to have two functions.  All ten need a particular processor.  The other function is different for each of the ten people.  If the ten other functions are in ten different chips, then you can easily build computers that each have two chips:  the processor and one other chip with the function that person wants.  If you want a single chip for each person, then you need 10 different chips, one for each combination, which is incredibly expensive to build.  Alternatively, you could have one enormous chip with all 11 functions, but then you've got an enormous die, which is still a problem.

    In practice, companies do integrate some functions with niche uses into their chips if the die space it takes is very small.  AMD's Tahiti chip has some hardware built in for double-precision (i.e., 64-bit) floating point computations that are useful for GPGPU purposes, but useless for gaming.  Nvidia's GF100 and GF110 chips and AMD's Cypress and Cayman chips did likewise.  The small amount of extra die space it takes is cheaper than building two separate chips.

    This also gives you additional binning options, where if you have a problem with double-precision not working right on a lot of chips, you can cripple it in GeForce and Radeon cards that you sell to gamers while picking the chips where double-precision works flawlessly for Tesla and FirePro (AMD discontinued FireStream) cards.  You can't disable it entirely on the gaming cards, as the OpenGL 4 specification (likely also DirectX 11, though I'm not sure here) says that GPUs have to be able to do double precision computations.  But OpenGL compliance only requires getting the correct answer eventually, and doesn't say you have to get it fast; that's why lower end cards tend to offer double-precision computations with only 1/24 to 1/16 of the speed of the single-precision (i.e., 32-bit) computations that are used so extensively in games.

    So yes, integrating more functions into the same chips as processors and graphics is the future.  Furthermore, processors and graphics in the same chip will be increasingly common as conditions allow.  Die size and power consumption are still problems at the high end, and memory bandwidth is a problem from the lower mid-range on up.  Those problems may well ensure that there are always discrete video cards for gaming.  But integrated graphics will be increasingly viable as time passes.

  • CabalocCabaloc Fort Pierce, FLPosts: 116Member
    Quizz, just wanted to let you know, yes I did take the time read all that .
Sign In or Register to comment.