Next-gen Radeons to be multi-core?

Modern graphics processors rely on extensive amounts of parallelism to get calculations done as quickly as possible, but those GPUs are still monolithic chips designed with a given number of stream processors, texture units, render back-ends, and the like, each depending on their performance grade. However, the folks at Fudzilla say they have it on good authority that AMD will challenge that paradigm with its next-generation graphics processor, code-named R700. According to Fudzilla, low-end, mid-range, and high-end R700 cards will all have GPUs with varying numbers of identical R700 cores. To determine speed grades, AMD will simply outfit higher-end cards with more R700 cores.

AMD’s top-of-the-line R700 product, for instance, will supposedly have four or more R700 cores in one die and will be able to crunch nearly two trillion floating point operations per second, or teraFLOPS. By contrast, Fudzilla explains that the existing Radeon HD 3870 is in the 500 gigaFLOPS range. FLOPS don’t tell the whole story, of course, but high-end R700 cards will be an order of magnitude faster than existing products if those numbers are even remotely accurate.

Interestingly, this rumor sounds similar to information that recently seeped out about Intel’s Larrabee project. Larrabee is expected to be a discrete, game-worthy Intel graphics processor scheduled for the not-too-distant future, and an Intel presentation nabbed by Beyond3D in April suggests Larrabee products will be based on multiple, small "throughput cores." A diagram showed a chip based on ten of those cores with a shared pool of 4MB of cache.

Comments closed
    • divided421
    • 12 years ago

    Ultra high-end R700 to finally play Crysis at 60fps in 2010!!!

    • l33t-g4m3r
    • 12 years ago

    for 3dfx reference: voodoo5 like=bad, voodoo2 like=good.

    With today’s modern technology moving more and more toward multiple processor designs, I don’t think this will be a bad product at all.
    I think by now they know all the pro’s and con’s, and will make it work.

    CPU’s certainly are moving this direction, it was only a matter of time before GPU’s did the same.

    • K9-Cop
    • 12 years ago

    Eventually we’ll have specialized multi-core. So on a 10-core chip you might have 6 cores for CPU-like functions and 4 cores for GPU-like functions. Or you could go even more specialized and say one of those GPU cores is going to perform Physics-like functions. It all depends on whether there is a performance gain by specializing for a certain function, and whether that certain function is needed by enough users. Multi-core CPU’s, GPU’s and AMD’s on-board memory controller is just a start.

    • dragmor
    • 12 years ago

    This does fit with AMD’s whole modular CPU plans. Its already been said that the 1st Fusion will be a dual die setup so i expect that we will have one of these cores and a CPU die on the same package at first.

    As for the GPUs it seems like AMD has already planned ahead a little.
    1) There current chips use a ring bus memory controller for the different clusters, you could still do this with multiple dies on 1 package with a little work.
    2) The R670 have about half the memory bandwidth on the R600 but are generally better performers. AMD said that they optimised the memory use in the GPU.
    3) The render logic on the 360 is separate from the main die and gives free 4xAA on 720p images but using the edram. This could work on the 700 GPU, the data from all cores goes to the one render unit that applies the AA/AF. 32mb of edram should give you 4xAA on 1920×1080 for free (4xAA because thats mandatory in DX10.1).
    4) AMD is putting a lot of effort into CrossfireX, the work done in these drivers should really help any multiple die GPU drivers.
    5) It makes economic sense, AMD is really pushing the process technology for the GPUs. Normally you manufacture your highend GPUs on the old process, the few times the highend was manufactured on the new process it turned out bad (dustblowers). Since there would be no difference between AMDs low end and high end chips they could continually be on the newest process tech. Plus only manufacturing 1 chip has got to be cheaper and easier to optimise.

    • Meadows
    • 12 years ago

    Am I mistaken when I say you will be able to have 16 distinctly recognised videocards in a system? (4 cards, 4 cores on each)
    Soon to evolve Crossfire X into Crossfire More-Geeky-Letter.

    • swaaye
    • 12 years ago

    Sounds like another way to make GPU tech that can scale across market segments. Each R700 core can be simpler than one giant core and they can just tack them together for the high-end part. Hopefully it works better than SLI/Crossfire. If they are all on the same package or die, communication can be stupidly fast.

    • provoko
    • 12 years ago

    Finally. With more transistors than CPUs, it only makes sense to put more cores onto a GPU. But the problem is, its still crossfire (or SLI if Nvidia went along). It’ll most likely be cheaper, so thats good. And with the cores being able to communicate faster because of their distance to each other, crossfire/sli may achieve 100% performance.

    • PRIME1
    • 12 years ago

    l[

      • Krogoth
      • 12 years ago

      LMAO

      • albundy
      • 12 years ago

      act-ully iz pronouzce jebuz.

    • Rza79
    • 12 years ago

    Multi-core on one die has many advantages. One being lowering debugging time. Both R420 and R520 have proved how difficult it is to debut an error in a complex chip. This was the reason that both were posponed for so long (6 months or something).
    This way of designing a chip would leave you with debugging two smaller cores: one part that holds all shared logic and one part that holds the gpu logic. That gpu logic is then copied over and over.

      • flip-mode
      • 12 years ago

      Good point.

    • wierdo
    • 12 years ago

    Smaller cores often increase yields vs one big hunk of inter-related logic, so even on a single die it can matter. This is why there are lotsa upcoming multi-core designs in the GPU and CPU markets… the shrinking of processes is leading to allot of free die area to fill stuff in, and putting multiple core there may sometimes offer better return on investment vs building something more complex to fill in that space, diminishing returns for that complexity and all.

    • StashTheVampede
    • 12 years ago

    Isn’t the real answer to ATi doing this: do they really have a choice? Intel and AMD have been progressing to multi-core for years (along with other players in similar markets), GPUs simply haven’t caught up.

      • snowdog
      • 12 years ago

      GPUs are already massively parallel. Going multi-die (note not multi-core) is mainly a yield benefit as it is easier to produce smaller chips. Also you improve economy of scale as you can use the same chip at multiple different performance levels.

      These benefits are countered by the difficulty of interfacing the chips later. Unless this solution is software transparent, then it gets ugly fast if you actually need crossfire drivers to work with games. Xfire doesn’t always work.

      Now if ATI has a hardware level inter working and no xfire needed, then this might be a good idea

        • Perezoso
        • 12 years ago

        AMD has the technology to make this solution software transparent: cache coherent HyperTransport.

          • shank15217
          • 12 years ago

          hyper transport isn’t fast enough for gpu scaling. Those gpu’s probably need to talk to each other at well over 80 GB/s.

            • Perezoso
            • 12 years ago

            I think you are overestimating the bandwidth needed.

            • provoko
            • 12 years ago

            80gb/s!?!? PCI-express already has too much bandwidth. You can hardly see a performance increase from 8x to 16x.

    • Dposcorp
    • 12 years ago

    If this runs as fast as it sounds, then I think a lot of the R&D that went into Phenom is being put to good use, and AMD just MAY end up righting the ship.

    If it is not supported well by drivers and newer games, then it may be the straw that broke the camels back.

    • Krogoth
    • 12 years ago

    I am not really that surprised.

    It is just becoming too damm difficult and economically impractical to keep designing and engineer more mega-GPU dies.

    In order words, the benefits of die shrinkage are shrinking while the penalties for leakage and increasing transistor budget remain.

    Nvidia will run into the same problem and opt for AMD’s more realistic path. Make a bunch of smaller, but separate dies and intelligently balance the workload. The difficult part and engineering challenge is the latter.

    • ReAp3r-G
    • 12 years ago

    would i be wrong to assume that when they mention multi-core graphics, its like slapping on an Athlon X2 on a graphics card? coz i’m pretty fuzzy on GPU architecture

      • d2brothe
      • 12 years ago

      That would indeed be wrong. This is more akin to slapping two radeon chips on one card, only…like an X2, they’re combined into one chip instead of having two cpu sockets.

    • Hattig
    • 12 years ago

    I assume that these cores would be connected via some form of crossbar, and that there would also be another “core” that incorporated all the shared logic, memory controller, video decode engine, etc.

    I don’t see how this is radically different from the current setup of varying the amount of of units in current GPUs.

    I could see it if the dies were separate, connected via a very fast form of HyperTransport for example. You’d get the benefit of smaller dies which have higher yield, and of matching dies that match a certain speed, and of scaling too (one die up to say 4 dies).

    (This is what AMD should have done in 2006 to get a Quad-core CPU out before Intel, but that company seems to be seriously lacking vision and is also putting idealism (native quad-core!!!) before making money.)

      • ReAp3r-G
      • 12 years ago

      well hopefully that honest-to-god idealism pays off in the end, they spend so much of money on it, it will pay off…i’m interested in AMD’s procs and stuff just for the pure fact of comparison and i’d hate to lose something to compare to…it will pay off i’m sure of it (maybe not now, but someday)

      • slaimus
      • 12 years ago

      3dfx Spectre has just this architecture.

      It had one chip that had all of the shared logic like AGP, T&L, and control logic, with 1 to 4 raster chips conneced to it.

    • Lord.Blue
    • 12 years ago

    Quad Crossfire on one card?

      • Nitrodist
      • 12 years ago

      oct crossfire with two cards ๐Ÿ˜›

        • poulpy
        • 12 years ago

        Will anyone dare mention 4 cards?

          • willyolio
          • 12 years ago

          hexdecafire. nVidia won’t touch that.

            • provoko
            • 12 years ago

            Fudzilla says up to *[

      • mortifiedPenguin
      • 12 years ago

      I suspect that this architecture, if true, does not mean using crossfire or a multi gpu driver based setup. I believe the big thing about R700 being a “multi-core” GPU has different implications than a “multi-core” cpu, load should be split natively between these two cores rather than rely on a driver to split the workload (which may or may not work). if it was just crossfire on a board, thats nothing special. they already have something like that, albeit not officially at the moment.

      in short, this shouldn’t be about crossfire or sli “on a chip,” it should be something completely new

        • provoko
        • 12 years ago

        That would be amazing. But it most likely will an on die crossfire.

    • flip-mode
    • 12 years ago

    l[

    • gtoulouzas
    • 12 years ago

    We’ve seen this ‘top-to-bottom’ multicore approach before, and the results were not pretty. Remember 3dfx’s death knell?

      • ReAp3r-G
      • 12 years ago

      not really, i didn’t follow graphics tech that far back ๐Ÿ™

      • d2brothe
      • 12 years ago

      Yes, but 3dfx made other errors. Besides, this is a little bit different than what they did, theirs was more akin to the dual GPU cards, this is more of a dual core GPU, which has proven very effective in CPUs. Plus, this is a different world now…the technology is different..

      • snowdog
      • 12 years ago

      Yes, I sold my 3dfx stock when VSA architecture was revealed.

      Multi-chip is a challenge to do correctly. You get yield advantages into breaking it into multiple dies, you also get to use the dies in a larger range of products.

      But in order for this to really be successful you need have the load sharing done at the hardware level, not with crossfire drivers and memory must be treated contiguously as well. If you need 2G of physical memory to behave like 512MB of memory on a Quad Chip card, then the design is a bust.

      It will be interesting to see how this works out.

    • Jigar
    • 12 years ago

    This is good.. But just the one thing to be aware of, the higher chips card carry the hungrier the card will be.

    • Shining Arcanine
    • 12 years ago

    I do not see the point of multicore GPUs as GPUs are inherently parallel such that you could simply double the number of everything on a chip and get a better effect because of all of the redundant circuitry and programming that would be needed to make a dual core GPU work that will not be necessary in a single core GPU. The only thing that is being saved with a multicore GPU is development time, not die space and definitely not power consumption.

    Perhaps GPU development is becoming too difficult for AMD or AMD wants to reallocate chip development resources elsewhere.

      • IntelMole
      • 12 years ago

      Double your die area and your yields will go down the pan.

      Break it up, however, and you can mitigate some of this effect …

        • MrJP
        • 12 years ago

        Yes, but they seem to not be talking about separate dies for each core, so doubling the number of cores does double the die area.

        In concept this doesn’t seem to be much different to what both Nvidia and AMD already do with different numbers of computational “units” enabled for different products. The only difference seems to be that the disabled units (or cores in this case) won’t physically be there on the low-end cards. Compared with an equal-performance monolithic design, the multi-core approach will presumably have the same yields for the high end part (or perhaps a bit worse given the extra logic needed for the cores to communicate), but should be better for the lower-end parts: smaller die (and better yields) than a partially-disabled high-end part and less design effort than a dedicated chip.

      • Flying Fox
      • 12 years ago

      Who said you need “redundant programming”, whatever that means? We still don’t know enough about this and they could have put in a “master controller” and have just one interface to the rest of the system, so you still address the thing as “one GPU” but on the board it dishes out the tasks to the cores.

      Of course we will have to see more details first to further comment on this.

        • Shining Arcanine
        • 12 years ago

        I guess redundant was not as good as a word as extra, as is needed when programming for multicore/multiprocessor systems.

    • Chryx
    • 12 years ago

    In the same way that a Radeon 9800XT is a dualcore Radeon 9600 Pro then….

    doubling the number of pipelines on a chip does not a dualcore make.

      • Forge
      • 12 years ago

      This is more like two 9600XTs on one die. The 9800XT was more integrated than that.

      Think shader core, plus ROP’ish core, plus memory controller core. Simply vary the number for different builds.

      It’s actually not too far from how NV/ATI do their in-line differentiation today. The 8800GTS was just 6/8ths of a 8800GTX. The only difference is that these future arches won’t be forced to tie everything from mem controller down to ROP into one package. They can disable/add parts from any part of the pipeline.

      Sounds good to me, and should be much friendlier to GPGPU programming.

    • Anomymous Gerbil
    • 12 years ago

    It had to be Cyril! Two to four times faster (2Tflops versus 500Gflops) is not what is generally referred to as “an order of magnitude” difference. Then again, we are in a binary world :-p

      • Xenolith
      • 12 years ago

      If my math is right, a single core R700 will be just as fast as a single R670 with floating point operations.

      • Gungir
      • 12 years ago

      Well, technically, 10 is an order of magnitude more than 9. ๐Ÿ˜›

      It may not always mean “ten times greater”, but you have to admit, it sounds damn good.

        • d2brothe
        • 12 years ago
        • kuwaiki
        • 12 years ago

        100 is an order of magnitude greater than 10, as 10 is to 1. Order of magnitude = x10

        You will learn it and use it (a lot) if u ever study any engineering…

          • DASQ
          • 12 years ago

          It’s a bit different if you were measuring on a Richter scale. Or… in decibels.

      • Xenolith
      • 12 years ago

      The proper term would of been, a /[

    • blastdoor
    • 12 years ago

    So, “SLI on a chip” ?

    I guess this might be better from a manufacturing standpoint than making bigger cores — higher yields for smaller chips.

Pin It on Pinterest

Share This