Comments Locked

133 Comments

Back to Article

  • imaheadcase - Wednesday, June 17, 2015 - link

    Do you happen to have a side by side to compare the size of this memory vs "old"?
  • DanNeely - Wednesday, June 17, 2015 - link

    Arstechnica has one. I'm not sure which card is the traditional PCB layout; but there hasn't been much variation in the PCB layouts of large single GPU cards for the last half dozen+ generations.

    http://cdn.arstechnica.net/wp-content/uploads/site...
  • phoenix_rizzen - Wednesday, June 17, 2015 - link

    The bottom one, with all the memory chips around the GPU, is the "old" layout.

    The top one is the new layout with the HBM memory on the GPU die.

    Not the greatest picture, but you can at least get a sense of the differences in the layouts.
  • Samus - Thursday, June 18, 2015 - link

    them is crazy looking memory "blocks"
  • Xpl1c1t - Thursday, June 18, 2015 - link

    the memory is on-die, HBM, coming to most gpus next year
  • FlushedBubblyJock - Tuesday, June 23, 2015 - link

    AMD must feel so stupid, their top tier cards have only 4GB ram, so they rebranded the 290X and 290 and overclocked them a tiny bit and put 8gigs on both because 2nd tier cards should have 8GB and top tier cards should have 4...

    ROFL - Man it must be so embarrassing to work there.
  • jordanclock - Tuesday, June 23, 2015 - link

    They only have 4GBs of RAM because that is a limitation in the current design of HBM. The idea is that while the capacity is lower, the outrageous amounts of bandwidth and power savings will be an offset.
  • WinterCharm - Sunday, June 28, 2015 - link

    Wont this be limiting at resolutions above 4K though?
  • MobiusPizza - Tuesday, June 30, 2015 - link

    In benchmark we see, not at the moment. Techpowerup writes:
    "My numbers show that at this time, there is no need for more than 4 GB of VRAM when targeting playable framerates. "
    Good thing is, memory usage is optimizable by game devs and drivers, we are currently very wasteful in game memory usage. Like Nvidia, AMD introduced new texture compression algorithmns which should alleviate maxing out the VRAM.
  • piiman - Saturday, June 27, 2015 - link

    " Man it must be so embarrassing to work there."

    It must be embarrassing to not know what you're talking about also.
  • Nate0007 - Tuesday, June 30, 2015 - link

    LOL.
    You took the words right out of my mouth. He really needs to grow up or go back to school and coninue his education. That or we should all take his word for it that he knows more about Graphics cards then the engineers at AMD. LMAO Poor guy, I hope he is not another Generation X living in his parents Basement.
  • Taristin - Wednesday, June 17, 2015 - link

    An interesting lack of capacitors on these new AMD cards. Were they needed for the old style VRam, or have AMD designed a new way of not needing them?
  • looncraz - Wednesday, June 17, 2015 - link

    I believe, if anything, they are using surface mount multilayer ceramic caps in an effort to keep the height profile down to minimize the amount of work that needed to be done on the heatsink/waterblock.
  • DanNeely - Wednesday, June 17, 2015 - link

    That's possible. Looking closely at new vs old single GPU boards, there is at least one component on the new one present in about the same area/quantities as 2 of the 3 sets of capacitors on the old board. They're low profile black rectangles marked either "159 C40" or "160 C40" and are polarity marked. There're 4 159 parts on the rear of the board (near the video out) vs 6 small electrolytic caps, and 14 160 parts on the front of the board vs 10 large electrolytic caps. I don't know how to interpret the markings on the 150/160 C40 parts; and don't have any pictures of the electrolytic caps in enough detail to read their capacity to tell if the apparent replacements are similar in size to the old caps or not.

    The more interesting bit is that unless their replacements are some of the 160's at the front of the board though, the 10 medium size electrolytic caps between the VRMs and ram are gone without a replacement. If these were used to make the ram happy I could see them being unneeded now; but if that was the case would've expected them to be spaced around the ram itself instead of all in a single area.
  • MrTeal - Thursday, June 18, 2015 - link

    The black caps are 16V rated 150uF Panasonic polymer tantalum caps that are on the 12V side of each power phase. The 150/160 is the capacitance in uF, Cxx gives the voltage rating (16V for C) and a lot code (the xx). The output caps would generally be between the inductors and the chip, but they're obviously not in this picture. I would imagine they're on the rear of the board.
  • CloseEnough - Wednesday, June 17, 2015 - link

    A close enough fix to perspective in the above image.
    As we can see there's a big difference in size

    http://www.mediafire.com/view/zk6uorln5ltmx9h/fiji...
  • meacupla - Wednesday, June 17, 2015 - link

    That's small. Possibly even small enough to fit a high power HBM GPU onto a half height/low profile card.
  • TeXWiller - Wednesday, June 17, 2015 - link

    They did have the R9 Nano concept card in the event. Probably the minimum size of the thermal solution is the limiting factor here.
  • Gigaplex - Wednesday, June 17, 2015 - link

    The Nano is full height, not half height.
  • HaryHr - Wednesday, June 17, 2015 - link

    AMD Radeon Fury MAXX seams like reasonable name.
  • Kevin G - Wednesday, June 17, 2015 - link

    No based upon horror that was the Rage Fury Maxx. Products that inspire nightmares should not be reborn.
  • Flunk - Wednesday, June 17, 2015 - link

    Crossfire was based on the Rage Fury Maxx. It pioneered alternate frame rendering.
  • PaulMack - Wednesday, June 24, 2015 - link

    Yeah, I still see no benefit of AFR. Are all dual GPU cards still AFR? I remember someone telling me you could use different methods, but I haven't seen this anywhere, or benchmarked.
  • Thatguy97 - Tuesday, June 30, 2015 - link

    The horror was really through drivers more than anything
  • Grayswean - Thursday, June 18, 2015 - link

    AMD Radeon Mad MAXX Fury Road.
  • DERSS - Thursday, June 18, 2015 - link

    AMD Radeon Mad MAXX Fury Road R9 Fiju Ultimate X2
  • Jtaylor1986 - Wednesday, June 17, 2015 - link

    Not sure how this isn't going to be crippled by TDP. If one FIJI alone uses 300W and the max power delivery is 375W for 2, 8 pin connectors + PCI Express socket.
  • ratte - Wednesday, June 17, 2015 - link

    read the Anandtech review of the 295x2 (500w) and you will find out why it won't be crippled by TDB
    http://www.anandtech.com/show/7930/the-amd-radeon-...
  • Andrei Frumusanu - Wednesday, June 17, 2015 - link

    Power consumption scales quadratically with voltage which is tied to frequency, having a both chips run slower than a single Fiji card should bring enough headroom and still give a great performance boost at a reasonable TDP.
  • dragonsqrrl - Saturday, June 20, 2015 - link

    "reasonable"
  • obsidian24776 - Wednesday, June 17, 2015 - link

    1x fury x was rated as 275w with 500w aio lcs
  • wiak - Wednesday, June 17, 2015 - link

    you forgot the HBM saves power part too :P
  • Gigaplex - Wednesday, June 17, 2015 - link

    And yet the Fury X still is a 300W card.
  • testbug00 - Wednesday, June 17, 2015 - link

    it's 275 watts average board power. I'm guessing it and the Titan X will be about the same performance once you give the Titan X the 110% power rating. The 980ti a few percent behind both at the 110% setting.

    AIB cards, everything's up in the air.
  • dragonsqrrl - Saturday, June 20, 2015 - link

    No, it's a 275W TDP, and based on the real world power consumption of the 390X, I wouldn't be surprised if it's avg power consumption at load also exceeds that by quite a bit.
  • lilmoe - Wednesday, June 17, 2015 - link

    "Mid-range" HBM interposer chips and SoCs is where its at IMHO, not really interested in niche "high-end"...

    I'll pass until I see a Zen APU with HBM (integrated 8-16GB DDR4), should be interesting to see the benefits of both power consumption and performance, and space area savings of course.
  • DanNeely - Wednesday, June 17, 2015 - link

    HBM isn't well suited as a replacement for DDR3/4 on the CPU. It's optimized for throughput (since in 3d rendering the GPU is mostly doing sequential reads of data from ram) not latency which matters for the more random IO of a general purpose CPU.

    If they can get costs down enough a 1GB HBM stack might work as on package VRam to help free the APUs bandwidth bottleneck problems. Replacing the DDR dimms used by the CPU for main memory will probably end up falling to a competing next generation memory system that's better suited for the CPUs needs.
  • tcube - Thursday, June 18, 2015 - link

    AFAIK HBM has much better latency then GDDR5 and quite possibly in the range of the DDR3-DDR4. HBM 2 could improve on that even more. HBM is most definitely good for at least normal PCs, it should make a great APU memory, especially so on laptops. Imagine a quad core zen with 1000SP's and 16GB of HBM in a 20-25W package using on 14nm FF process? That thing would be absolutely insane! Much more power then an average user would need on a day to day basis. And could run any game in hd(max settings)-fhd(high settings), should be able to even run some VR.
  • geekfool - Friday, June 19, 2015 - link

    "HBM isn't well suited as a replacement for DDR3/4 on the CPU."
    that's what WideIO2 is for....
  • Lolimaster - Thursday, June 18, 2015 - link

    The APU only needs 2GB of HBM memory, then you can go cheapo 2133 DDR4 memory for the system.
  • TomWomack - Wednesday, June 17, 2015 - link

    Why are there only four HBM stacks? It looks as if you could fit three down each side of that ludicrously big GPU die without enlarging the even-bigger silicon imposer.
  • lilmoe - Wednesday, June 17, 2015 - link

    Give it more time to mature. Fitting "only" 4 stacked modules isn't as easy as it seams.
  • Mark_gb - Wednesday, June 17, 2015 - link

    This is the first time any video card company has worked with HBM. They were putting the pieces into place as much as 2 years ago. And trying to figure out what they could do today. You can almost bet that they thought they would be doing at least 16nm by now, if not 14nm. And as we all know, both AMD and Nvidia have been stuck on 28nm.

    HBM will move to smaller nodes, stack higher, run at a faster speed, gain DDR technology, and all of that in a relative short period of time. And they should have enough confidence by next year to start using more HBM stacks with a GPU. Everything new always starts big and slow... and shrinks quickly as engineers understand it more, and become more comfortable with the product.
  • MrSpadge - Wednesday, June 17, 2015 - link

    Plus:
    - they don't need more bandwidth than the 0.5 TB/s they already get now.
    - and the chip is already so large at 596 mm², with 600 mm² being the current practical limit, that there was no space left to fit even one more of the comparably small HMB controllers
  • Kevin G - Wednesday, June 17, 2015 - link

    Largest commercial die produced was the Tukwilia Itanium at an impressive 699 mm^2. Theoretical limit for how large a die can be is between 700 and 800 mm^2 during that era.

    If foundries ever move to 450 mm wafer sizes, producing such large chips will become far more feasible. However with interposers I would predict a reversing of that trend to smaller dies. Far easier to get good yields on two smaller dies than one larger unit. It makes sense from both a performance and cost perspective for designs using an interposer for HBM anyway.
  • tcube - Thursday, June 18, 2015 - link

    Yeah! I would love 2 fiji dies on an interposer with a 4x HT(4.0?) link between the two, should make them work as one single gpu... imagine putting multiple smaller chips all interlinked with ht links, say use 1 link/chip and use smaller chips(1k sp's each) and a larger interposer. Would make the package quite complex but you should get much better scaling then with the cfx bull
  • Kevin G - Thursday, June 18, 2015 - link

    It wouldn't even be HyperTransport, it'd be the native internal bus used on-die. Not need to use a topology like HT that was designed to work between sockets.
  • tcube - Thursday, June 18, 2015 - link

    Where did you read that fiji has a 596mm^2 diesize?(I knew something along the lines of 550-568) Btw Titanx has a diesize of 603^mm
  • Kevin G - Wednesday, June 17, 2015 - link

    A bit of a nitpick but HBM already uses DDR bus. Most places are reporting a 1 Ghz clock speed but that's the effective rate. The real clock speed is a meager 500 Mhz.

    HBM2 is due next year and will bring both capacity and bandwidth increases on a per stack basis. As mentioned, there is room for more stacks if need be as well.

    Though I'm of the mindset that instead of increasing die size and number of HBM stacks per die to astronomical levels, AMD will return to a small die strategy and simply use an interposer for in-package Crossfire. Bandwidth between dies shouldn't be an issue and latencies between units would be similar if the chip was monolithic. Say four GPU dies + eight HBM stacks for the high end product, two GPU + four HBM stacks for a midrange product and a single GPU die + two HBM stacks for the low end. Their SoC efforts could also benefit by having a dedicated CPU die and simply re-use this GPU die to complete the design. Rinse and repeat for the proper mixture of CPU cores and GPU assist. AMD would have a full product line up using only two actual logic dies plus a range of interposers. That would greatly simplify development, something AMD needs to be focused upon.
  • K_Space - Wednesday, June 17, 2015 - link

    Will Pascal launch with HBM2 then?
  • DanNeely - Wednesday, June 17, 2015 - link

    Yes.
  • Mr Perfect - Wednesday, June 17, 2015 - link

    Sounds cool, but wouldn't they have to do alternate frame rendering on everything then? Multi-GPU doesn't scale as well vs a monolithic die, and not at all if the drivers or game aren't configured properly. Imagine buying a $700 card and having it perform the same as the $100 one because it's only effectively using one of the GPUs. The sales would be dismal. DX12 would sort that all out of course, but seeing as developers are still hesitant to force everyone to DX11, that could take another five+ years.
  • Kevin G - Wednesday, June 17, 2015 - link

    The thing with an interposer is that you don't have to limit the design to narrow buses that have to traverse relatively great distances. HBM is a prime example example of this as the 'narrow' 512 bit wide, 5.5 Ghz effective bus for GDDR5 on the R9 290X is being replaced by a 4096 bit wide, 1 Ghz effective bus. Similarly, being able to coordinate mulitple GPU's is not limited to 16x PCIe 3.0 bandwidth and topology. Want a wide, private bus to link the tessellation units together to act as one large virtual tessellation unit? That is now possible. Effectively by using to link multiple GPU dies together, it becomes feasible to link them in such a way to present them as one large virtual GPU instead of two or four independent chips. Performance wouldn't be perfectly linear but with the embarassingly parallel nature of graphics, it'd be pretty close.
  • DanNeely - Wednesday, June 17, 2015 - link

    I expect DX12 uptake to be much faster than DX10/11 have been. DX10/11 only offered major eye candy gains for people with top end GPUs. Among the gaming populace as a whole (vs Anandtech readers) there are far more people playing at 1080p medium on a $100 card, 720p low on an IGP, or even what was a medium high end card when new but that's no faster than new budget cards today; than at 1080p Ultra or higher on a newish $300+ card.

    Even at the lowest feature level, DX12 offers big performance improvements across across anything that supports it; and both AMD and nVidia are committed to bringing it to their last several generations of GPU. (Has Intel said how far back they intend to offer DX12 drivers?)

    I expect being able to game 'one quality level higher' for free will drive much faster uptake of Win10/DX12 than has been the case with meaningful DX10/11 (DX10/11 on a card that isn't fast enough to use any of the new eye candy doesn't matter) has been. While legacy versions of DX will probably continue to be available at the lowest quality settings for some time, two years from now I would be surprised if high/ultra settings aren't DX12 only (cheaper to only code the effect once) on most major titles. I wouldn't be surprised if the same is also true of medium quality on some as well.
  • tcube - Thursday, June 18, 2015 - link

    Yeah, a 1000SP GPU die, a quad CPU die and an uncore/io/nb die(as intel does with the pch) and mix and match with HBM stacks... but I would use HT(a new iteration if needed to pump up the throughput) links not cfx crp between the GPUs. If AMD would pull this off they would be able to build quite complex chips, and have distinct small teams working on distinct dies that just get plugged into the interposers as needed. They could scale up or down a GPU or CPU or reballance a chip in a matter of a month or so if competition throws them a curveball... They could even marry 28nm chips to 14nm chips and such... While at the same time have incredible yields on those chips.
  • FlushedBubblyJock - Tuesday, June 23, 2015 - link

    Maybe you, tcube, and Kevin G need to send your engineer dev credos to AMD so they get on with it, or better yet you two could be CO-CEO's and run the company correctly as soon as ET verifies your joint fantasies.
  • Aikouka - Wednesday, June 17, 2015 - link

    It's limited by the memory bus. You can toss as many memory chips on there as you desire, but unless you have any I/O to tie the traces to, they're quite literally useless.
  • Gigaplex - Wednesday, June 17, 2015 - link

    The large number of tracks required to connect the stacks is the limiting factor. Future products will use higher capacity stacks.
  • Lolimaster - Thursday, June 18, 2015 - link

    They will will ship an 8GB Fury X, same 4 stacks but with double the density (256 --> 512MB)
  • CPUGPUGURU - Wednesday, June 17, 2015 - link

    What good is saving space when it needs a massive space wasting radiator? 4GB of VRAM is too little to be considered 4K gaming future proof.

    28nm GPUs will never come close to using HBM bandwidth, HBM just isn't needed until the next process node 16nm, GDDR5 is not bottlenecking 28nm GPUs so 4GB HBM1 is a marketing gimmick that's 4K gaming insufficient.
  • jackstar7 - Wednesday, June 17, 2015 - link

    You're saying a lot of things with no data anywhere to support your assertions. Just so people don't confuse your conjecture for facts.
  • CPUGPUGURU - Wednesday, June 17, 2015 - link

    Get out of the cave you've been living in and OPEN your eyes to the Facts.

    The tech forums are full of gamer with mods using more than 4GB of VRAM but you are blind to those facts. You are also blind to the Fact that NO 28nm GPU in the world is memory bandwidth bottleneck with a properly designed GPU. YES, the next process node 16nm will need HBM, but not this many years old 28nm process.
  • DigitalFreak - Wednesday, June 17, 2015 - link

    Yeah, this "4GB is plenty of VRAM" crap is getting old. It's been proven multiple times that some games will use over 4GB, and not just ones that are modded. http://www.hardocp.com/article/2015/06/15/nvidia_g... is a perfect example.
  • CPUGPUGURU - Wednesday, June 17, 2015 - link

    For making a future proofing 4K gaming buying decision 6GB is a MUST HAVE, 4GB is too little and too lame to 4K game.
  • Black Obsidian - Wednesday, June 17, 2015 - link

    Given Fury's significantly increased memory bandwidth and (alleged) additional compression algorithms, we don't know if it's subject to exactly the same limits as other cards.

    Surely we're all adults who can handle waiting a week or two for concrete data before making blanket statements.
  • masouth - Wednesday, June 17, 2015 - link

    Yes, it will be nice to actually get the benchmarks but there is some validity to people's worries so far.

    If a game requires size X amount of textures to be held at certain resolutions and settings then bandwidth will mean absolutely squat. 4GB can't hold more than 4GB at a given time no matter how fast it is because speed has nothing to do with the capacity...just how fast you can move information into and out of that set capacity. =/
  • CPUGPUGURU - Wednesday, June 17, 2015 - link

    masouth knows his Facts,

    Bandwidth, "mean absolutely squat" if more than 4GB of VRAM is needed performance is killed and there are NO 28nm GPUs that can utilize all the bandwidth that HBM provides, other than a smaller form factor that's negated if you add a massive radiator and reducing wattage HBM1 is just not needed on 28nm GPUs. AMD rushed into HBM thinking we would have GPUs built on smaller process node with many more transistors and that didn't happen with 20nm. Next year GPUs using 16nm TSMC will need HBM2 with vastly more GB of memory available, having only 4GB is future proofing 4K gaming obsolete.
  • eachus - Thursday, June 18, 2015 - link

    Sigh! There are two questions here, and they interact a lot. First, how good is AMD's recent compression technology? The good/bad news can be summed up by the R9 285 card. You can buy 3 Gig 280 and 280X or 4 Gig 290 and 290X cards. I think I have even seen 8 Gig 290X cards floating around. But the card manufacturers, given the option (and it is a pretty easy option) to build 285 cards with 4 Gig of memory have, in effect refused to do so. Why, because of the compression the 285 can compete with the (3 Gig) 280 and 280X that are still available--at least for smaller displays.

    Once DX12 comes along, I expect that for 1920x1200 or smaller displays, a 2GB Tonga will be fine. For now buy a 280X or 290X if you are at 2560x1600 or so. I hope that 4 GB 380X cards will show up soon. (The R9 380 renames Tonga) but no 380X has been announced yet. I think AMD may be holding off to see the supply and demand for HBM. AFAIK, Tonga can use HBM but there is no point at 2 Gig.

    So what happens with Fiji. The simple answer is that 8 Gig cards will come (and 16 Gig dual GPU cards). But early next year is more likely than later this year. Practically speaking, though a 4 Gig 390X should use about 2/3rds of the memory needed for the nVidia 980 Ti. If these benchmarks are correct: http://www.techpowerup.com/213528/radeon-fury-x-ou... the picture is very clear... At 4K the (4 Gig) 390X outdoes the Titan X and 980 Ti. At 5K that is reversed, and at 8K resolution only the 8 Gig Nano and Titan X survive.

    When will we see 8 Gig 390Xs (and 4 Gig 380s)? Pretty soon I think. The issue is memory availability more than anything else.
  • Mark_gb - Wednesday, June 17, 2015 - link

    Not according to AMD. And since they designed this card, know what technologies are built into the silicon and worked into the drivers, I suspect they know far better than you and I. And they say the Fiji chip will handle all 4K gaming just fine.
  • Mark_gb - Wednesday, June 17, 2015 - link

    You cannot use an Nvidia card to make your point about a brand new GPU from AMD that was designed from the ground up to do 4K video.

    By your logic, a rocket exploded back in the 1950's... So all future rockets, no matter who makes them, will also explode.

    That doesn't work. Show us a real use situation on a Fiji processor where 4GB is not good enough, and then we'll say its not. But don't come walking in here with some fantasy situation like Prime95 generates on a CPU, and try to make us believe it.
  • CPUGPUGURU - Wednesday, June 17, 2015 - link

    You're assuming that back when AMD was designing its not new Tonga/Fiji that 4K gaming was on the map, it wasn't, 4K monitors were very expensive, no one had any idea 4K would be the future proofing standard it is today, even now its a year or two from prime time gaming.

    AMD is hiding this fact by keeping testers from benchmarking Fiji and pumping Fiji full of propaganda before it gets busted by the 4K gaming Fact that 4GB is too little and too lame to 4K game. When any memory Hits the 4GB limit its performance will be destroyed, and that's already a 4K gaming Fact that HBM1 can't escape.
  • Oxford Guy - Saturday, June 20, 2015 - link

    Funny how the craptastic VRAM of the 970 (28 GB/s for the second partition), especially in SLI, hasn't garnered nearly the same level of scorn. Nvidia is still advertising the card as offering 4 GB of 224 GB/s speed on their website, even.
  • FlushedBubblyJock - Tuesday, June 23, 2015 - link

    Oh really, you hammered it worse than any of the above, and so did and do all the amd fans.
    The problem and difference is it was already released and benched - benched a lot and found to be killer perf for price.

    In this amd case, we have a tech limit and everyone should be very concerned, as we have not seen the results other than sparse, specialized, amd leaked benches.
  • Mikemk - Wednesday, June 17, 2015 - link

    So would you rather they wait till the next node and release HBM as a new possibly buggy expensive technology, or release it now and have it be more reliable and slightly cheaper when it is needed?
  • CPUGPUGURU - Wednesday, June 17, 2015 - link

    You assume HBM would be buggy you assume wrong, and its expensive now, so your assuming points are pointless.
  • FlushedBubblyJock - Tuesday, June 23, 2015 - link

    I have to laugh Mikemk, because amd fans blabbered over and over how amd was cutting edge ddr5 technology and they have the upper hand releasing it and blah blah blah ... now amd ddr5 goes 5k or 6k and nvidia's is a solid 7k and oc's to 8k, thus meaning amd screwed it up and still is screwing it up in comparison.
  • xthetenth - Wednesday, June 17, 2015 - link

    I'm sorry to hear your computer doesn't have extra fan slots.

    A radiator's hardly wasting space when it's allowing more flexibility and lower temps compared to a card using a similar volume.
  • CPUGPUGURU - Wednesday, June 17, 2015 - link

    There is no sense to saving space where space is not a problem, I never had a space issue with the length of GPU cards but placing another very thick radiator is a major problem. Its a bad space saving joke when the card needs a massive radiator, how hard is that to understand.

    A air cooled card takes up less space than what you AMD pumpers are pumping and a air cooled is much easier to install. So quit pumping the small space saving BS when you need mount a massively thick radiator close to your marketing gimmick 4GB HBM1 that's 4K gaming insufficient and not needed on 28nm GPUs.
  • SonicKrunch - Wednesday, June 17, 2015 - link

    Have I missed the DX12 announcement change that they won't be able to effectively use SLI(as this card essentially is) Full VRAM? DX12 will allow both cards VRAM to be used now. So this will have 8GB of VRAM effectively under DX12.
  • Kjella - Wednesday, June 17, 2015 - link

    Not even close to having 8GB, more like the way the GTX970 has 4GB of RAM where 3.5GB is fast and 0.5GB slow only the last 4GB will be even slower and in use by another GPU. With alternate frame rendering like CF/SLI does today almost all assets are needed on both cards. You need to write a tiled rendering path so they can divide the work spatially and not temporally, but this has obvious problems with reflections, shadows and such that impact the whole frame. And with single cards giving GTX980 Ti/Fury performance, it won't be worth it. I expect that outside tech demos 95%+ of developers will use DX12 like they use DX11.
  • Urizane - Wednesday, June 17, 2015 - link

    Saying 4GB of VRAM on a card like this would be fast while the other 4GB would be slow is like saying NUMA doesn't work. There's plenty of existing application code that has dealt with utilizing separate pools of memory attached to different processors to complete the same task successfully. Essentially, all we need is a framework similar to NUMA that works on DX12 and we'll have/eat all of the 8GB cake.
  • extide - Wednesday, June 17, 2015 - link

    It doesnt work at all like NUMA -- the card's won't be accessing each other's memory -- the PCIe bus is just not fast enough. They will need all textures local just like they do now, so yeah while each card's memory is individually addressable -- most of the data will still have to be duplicated anyways.
  • sabrewings - Saturday, June 20, 2015 - link

    The caveat to this is that it requires developer implementation. DX12 is putting developers much closer to the actual silicon, and features like asynchronous GPUs will require developers to make use of it. That's why I've always felt dual GPU cards are gimmicky. It's a lot of money and power to run the risk of a lack of software support leaving you with effectively a single GPU.
  • JMC2000 - Wednesday, June 17, 2015 - link

    With my Antec 300, I can mount the radiator either at the back fan emplacement near the CPU, on the side panel fan area, above the CPU in the 140mm fan space, or at the front of the case with a bit of modding.

    Unless you have a smaller case, or just don't want to change fans, I don't see how placing the rad is a problem.
  • extide - Wednesday, June 17, 2015 - link

    Then buy an air cooled fury x DONE. WHy are you bitching about something that is a NON ISSUE?
  • YazX_ - Wednesday, June 17, 2015 - link

    There is nothing called future proof, things are moving extremely fast, these GPUs will dry out way before 4k becomes mainstream and playable on mid-range cards, for now 4GB Vram is enough, though 6GB is the sweet spot but 8GB is useless , and still we dont know how game devs are moving with DX12, if they used it properly then 2GB Vram will be plenty to load very high resolution textures with Tiled Resources.
  • Jtaylor1986 - Wednesday, June 17, 2015 - link

    And yet they though 8GB of ram was important enough to put on the 390 (X) even though it costs more when margins are thin and generates and power when they are at the limit with both.
  • SonicKrunch - Wednesday, June 17, 2015 - link

    DX12 will allow this card to use all 8GB on board...
  • CPUGPUGURU - Wednesday, June 17, 2015 - link

    The very definition of making the best future proofing buying decision with the tech available now is more than 4GB of VRAM for 4K gaming which AMD missed the boat on with their HBM1.

    NO 28nm GPU in the world needs all the bandwidth the HBM1 provides, HBM1 is a total waste of bandwidth on 28nm GPUs. BUT what 4K gaming does need is more than 4GB, high end cards need to have more than 4GB of VRAM to even be considered future proof.
  • extide - Wednesday, June 17, 2015 - link

    This GPU needs the b/w so .. your theory is incorrect -- and anyways the manufacturing process doesnt really have much to do with it. I mean clearly the engineers who did the math figured that 512GB a sec is required for 4096 shaders, and 512GB a sec would be pretty tough to get on GDDR5 -- I mean you would need a wide bus going really fast and that would be expensive, require complex pcb's with more layers and use more power. The HBM is a very elegant solution to all of that.
  • CPUGPUGURU - Wednesday, June 17, 2015 - link

    That's a AMD problem not a NVIDIA issue, NVIDIA has been blowing AMD wide bus marketing BS into the stone age for years now. NVIDIA has been using compression since Fermi and has out performed AMD with a narrower Bus for years now while AMD has been using the Big Bus for Marketing bullet points. HBM is elegant but with only 4GB its 4K gaming obsolete, HBM2 is the real HBM deal, wait for it, or buy a 6GB Maxwell 980Ti, thats how we are rolling.
  • EpicFlails - Wednesday, June 17, 2015 - link

    Yeah and some will disagree that they have been blowing it out of the water. The compression needed due the low bandwidth has crushed Nvidia image quality as evident here.
    http://www.hardocp.com/article/2015/05/26/grand_th...
  • CPUGPUGURU - Wednesday, June 17, 2015 - link

    FXAA Yes, MSAA/TXAA No, its not due to low bandwidth so quit flinging the FUD, its a option for those without high end hardware to optimize their graphical settings to the performance capabilities of their hardware. Not everyone is running a new top end GPU, the 980Ti wasn't available(May 26) to test so its not relevant.

    Therefore, some objects Softest shadows look better, some objects AMD CHS looks better and some can argue NVIDIA PCSS looks better/more realistic on some things, but not all.

    There doesn't seem to be one setting that does everything perfect. Each one has issues, AMD CHS suffers from some terrible noticeable dithering on the edges of shadows up close, in some situations. Softest doesn't define the shadows completely well. NVIDIA PCSS perhaps over softens shadows to the point you can't even see these in some cases, and in others creates jagged edges.

    Thanks for the link
  • Pantsu - Wednesday, June 17, 2015 - link

    Your precious 980 Ti will be just as obsolete after 16 nm cards come out. For any kind proper scenario where you want over 4 GB VRAM, you'll need more power than what these cards are able to offer.
  • CPUGPUGURU - Wednesday, June 17, 2015 - link

    Wow you're so smart for Stating the Obvious, so what's your point for making a buying decision in the here and now?

    Maxwell 6GB 980Ti is your best 4K future proofing gaming bet, that's my point. Get a zotac geforce gtx 980 ti amp! extreme or gigabyte geforce gtx 980 ti g1 they offer about 30% more performance that a stock clocked 980Ti, that's how we will be rolling.
  • FifaGamer - Wednesday, June 17, 2015 - link

    i don't hold onto any sides normally, but no matter how i look at it Fury looks like a better Engineered card at the moment. I will reserve by judgement on 4GB till i see the benchmarks and based on what my friends under NDA told me, initial testing its beating 980 Ti comfortably in most games. buying a 980Ti was a very good decision by you and its still a good card. no need to be butthurt, other people will make their buying decisions after looking at the benchmarks and not looking at hopeless rants by you
  • CPUGPUGURU - Wednesday, June 17, 2015 - link

    You just proved AMD is lying about drivers not being ready as the reason it hasn't let reviewers benchmark Fiji. AMD has lost all credibility and showing Fiji in a closed room full of paid to pump puppets that gullible you happen to know doesn't prove anything, it just shows how desperate AMD is to hide the benchmarks.

    The only fair way to benchmark Fiji and Maxwell 980Ti is air vs air water vs water cooling, so will have to wait for the benchmarks from a site like this one, because I sure don't believe what a debt laden desperate AMD pumps.
  • BillyHerrington - Wednesday, June 17, 2015 - link

    What the heck is wrong with you dude ?, does AMD kill your cat or something ?
    Anandtech badly need moderator, i think Tomshardware is better place for hardware discussions since troll from kid are always deleted or hidden.
  • FlushedBubblyJock - Tuesday, June 23, 2015 - link

    He made good points, and calling names and insulting is nothing to be proud of billy h, just because you didn't like his solid logic because amd was in a bad light with it you need moderation
  • Nate0007 - Tuesday, June 30, 2015 - link

    How old are you ? I mean really.
  • loguerto - Thursday, June 18, 2015 - link

    (•_•) .. <-----------------Nvidia
    ∫\ \___( •_•) <---------------------CPUGPUGURU
    _∫∫ _∫∫ɯ \ \

    ( ͡° ͜ʖ ͡°)
  • loguerto - Thursday, June 18, 2015 - link

    (•_•) .. <-----------------Nvidia
    ∫\ \___( •_•) <---------------------CPUGPUGURU
    _∫∫ _∫∫ɯ \ \

    ( ͡° ͜ʖ ͡°)
  • FlushedBubblyJock - Tuesday, June 23, 2015 - link

    6GB is already thin, so yes at least 8 is needed, but the 390X just doesn't have the core power to justify 8GB.
  • Nagorak - Thursday, June 18, 2015 - link

    Future proofing is kind of pointless on a high end GPU. Yeah, it sounds weird, but I'd argue it's true. The people who buy these halo products are the same one that upgrade to the latest and greatest whenever something new is released. They won't be using the card in a year and a half.

    If you only upgrade every couple years and need to be concerned about future proofing, then you shouldn't be buying the super high end stuff anyway.
  • sabrewings - Saturday, June 20, 2015 - link

    That's a wide generalization. I don't care about future proofing, and I don't buy the latest and greatest every year. I bought a 980 Ti because its aligned with when I was building my PC. My last PC was put together in early 2008 and I was still running my Q6600 and GTX 275 (playing Elite, no less) up until this month. I buy the best to make it last, but I don't expect to play with settings maxed out forever either. I have a 55" 1080p LED TV (for which my liquid cooled 980 Ti barely hits 23% utilization and is usually around 26C in Crysis), but the real reason I bought the 980 Ti is for VR. StarVR, SteamVR, Oculus, and even a Hololens will all eventually grace my machine and I wanted it to keep up without needing more upgrades.

    All that being said, I'm not opposed to offloading my 980 Ti late next year or keep it as a spare and rocking Pascal since it probably will be a huge leap forward with such a large node jump (nearly 50% feature size reduction).
  • tcube - Thursday, June 18, 2015 - link

    HBM was used on 28nm not because of the bw... but because of the saved space on the GPU die itself as the interfaces are a factor smaller then GDDR5, also they did it to reduce power requirements of the entire package. The HBM stacks use less power so do the interfaces, less heat as well. I'd say it is a good move as it enabled them to put the power saved against the cores and increase the core count by more then 50% over Hawaii but use LESS power. And no you don't NEED a radiator you could use an aircooling solution but they overengineered the board to be a premium/enthusiast part... you don't get that fine, buy the aircooled version! The cooling solution is a 500W monster... the GPU under it apparently uses about 200W in games and has a tdp of 275W... that means you have a huge headroom, also you can deliver 375W to the board but also not needed... as I said it is engineered to be a premium device! You don't like it... get the aircooled version.
  • sabrewings - Saturday, June 20, 2015 - link

    I'm not so sure I believe that their puny AIO cooler is going to acceptably displace 500w (at least to my standards). An XSPC RX360 barely pushes 500w at 10* delta-T (which is the upper limit IMO for coolant temps). With it my 980 Ti load temp is around 36*C. I could probably toss a second one in and see around 40C load temps, but that's as high as I'd prefer to go. If you're really expecting that cooler to move 500w, you're going to have a huge delta-T which shortens pump life as well as significantly raising your silicon temps.
  • FlushedBubblyJock - Tuesday, June 23, 2015 - link

    "probably uses about 200W in games"
    ROFLMAO
    another one, they're everywhere
  • loguerto - Thursday, June 18, 2015 - link

    (•_•) .. <-----------------Nvidia
    ∫\ \___( •_•) <---------------------CPUGPUGURU
    _∫∫ _∫∫ɯ \ \

    ( ͡° ͜ʖ ͡°)
  • hingsun - Wednesday, June 17, 2015 - link

    WITH STACK HBM RAM, I AM REALLY WORRY ABOUT HEAT ISSUE, SINCE NORMAL CHIPS RECEIVED MORE SURFACE AREA TO DISSIPATE HEAT, WHILE THE 2.5D STACKS HAS MUCH LESS SURFACE AREA TO VOLUME RATIO TO DISSIPATE HEAT. WE ARE TRADING IN CARD SIZE VS CHIP LIFE. AS FAR AS I KNOW, MORE HEAT TO CHIP, THE LESSER ITS LIFE.
  • Mikemk - Wednesday, June 17, 2015 - link

    The button between your left shift and tab will allow you to type normally.
  • Black Obsidian - Wednesday, June 17, 2015 - link

    Also, power requirements for HBM are significantly lower than with GDDR5. So less surface area for cooling, but less heat that needs to be dissipated, too.
  • extide - Wednesday, June 17, 2015 - link

    The GPU is fully in contact with the heatsinking device just like it always has -- the ram chips do not cover up the gpu so this is not an issue.
  • extide - Wednesday, June 17, 2015 - link

    I see you're talking about the ram dies itself -- well ram doesnt get that hot anyways and this ram is clocked lower and runs on lower voltages so less heat so again, its not an issue.
  • FlushedBubblyJock - Tuesday, June 23, 2015 - link

    not an issue IS SITTING RIGHT NEXT TO THE GPU CORE TOP BLASTING IT'S HEAT RIGHT NEXT TO IT.
    Other DDR5 is off die and aoff substrate and off in the wind of the fan off on another part of the card...

    SO YOU'RE STATEMENT IS BAKED.
  • goldstone77 - Thursday, June 18, 2015 - link

    HMB creates less heat though.
  • FlushedBubblyJock - Tuesday, June 23, 2015 - link

    But it creates it heat right next to the gpu core mm's away slamming the gpu

    Unlike DDR5 which is miles away from the gpu core in comparison

    HBM is smoking the gpu core up as it sits right on it breathing down it's neck on all sides
  • Clauzii - Tuesday, June 30, 2015 - link

    It's the GPU that heats up the HBMs, if anything. But the cooling solution nulifies any heat-problems overall.
  • zepi - Wednesday, June 17, 2015 - link

    It would be interesting to see dual-gpu on a single interposer with some way increased bandwidth between them allowing DMA between memories etc.

    This should allow way better xfire than pci-e bridging, and i'd guess especially dual-screen VR could benefit enormously.

    Or maybe even CPU & GPU & shared memory on same interposer. True HSA, not the pci-e or ddr3 starved abominations we have now.

    And cooling with water.
  • JMC2000 - Wednesday, June 17, 2015 - link

    They would have to use linked interposers, if the GPUs don't become extremely small. But, it would still need a PCI-e bridge, unless AMD was to go back to the X1900 way of Crossfire with a 'master' and 'slave' GPU.
  • Wirmish - Wednesday, June 17, 2015 - link

    AMD Radeon Fury Road, with the new Fiji die.

    "Oh, what a die! What a lovely die!"
  • DroidTomTom - Wednesday, June 17, 2015 - link

    By nature won't this card have the 8GB of RAM everyone is concerned about? Since each die has 4GB of HBM for 2x4GB=8GB.
  • Pantsu - Wednesday, June 17, 2015 - link

    It has 8 GB of VRAM total, but when using AFR CrossFire it'll have to have the same data copied for both GPUs so you still have only 4 GB for buffers and assets which are copied to both GPUs. DX12 will give the option to do SFR CrossFire which could technically utilize the full 8 GB without doubling the data. So far we've only seen this in the latest CIV using Mantle API. It's not guaranteed games will shift to using SFR with DX12 though, and even if they do, it's not an instant transition by any stretch. It requires more work for devs, and they're pretty lazy as it is to provide multiGPU support.
  • DroidTomTom - Sunday, June 21, 2015 - link

    Thanks! I see what you are saying. It will be interesting to see the benches for this in DX11, DX12, and Mantle games.
  • goldstone77 - Thursday, June 18, 2015 - link

    AMD knows you need 8gb for 4k they aren't stupid. Don't know why people are saying 4gb.
  • FlushedBubblyJock - Tuesday, June 23, 2015 - link

    How dare you !
    AMD has gone where no card has gone before !

    I called them Daddy Dragon when I ran that platform, and The Real Spiderman when I ran that !

    Groundbreaking bulldozing piledriving technological advancement is not to be belittled !
  • CPUGPUGURU - Thursday, June 18, 2015 - link

    This dual GPU card uses seperate interposers, AMD took to easy way out, they needed to make a one interposer that can handle 2 GPUs.

    But this Dual GPU card attached to massive Dual Fan radiator will all the same crossfire problems (drivers, scaling, micro stuttering) that have plagued AMD FOREVER.

    Dual GPU cards have next to no market share because you're much better off buying one air cooled card now and when prices drop as they always do buy another.
  • 3ogdy - Wednesday, June 17, 2015 - link

    "So, while the performance advantage of the Rage Fury MAXX's 4.6GB/s peak memory bandwidth over the GeForce SDR's 2.7GB/s won't be seen running at 640x480 at 16-bit color, the MAXX will begin to pull away from its NVIDIA-born counterpart at resolutions of 1024x768 at 32-bit and at higher resolutions."

    http://www.anandtech.com/show/438/2
  • goldstone77 - Thursday, June 18, 2015 - link

    that's common sense, look at the clock speeds vs. ram.
  • goldstone77 - Thursday, June 18, 2015 - link

    AMD is aiming for 4k screens.
  • Lolimaster - Thursday, June 18, 2015 - link

    I would probably get the Ruby Nano and apply a nice underclock/underclock (say 800Mhz for the gpu). Since I won't do 4K the gpu will have a ton of unused power, better go eco.
  • Lolimaster - Thursday, June 18, 2015 - link

    Underclock/undervolt

    *Where is the edit button?
  • Bigman397 - Wednesday, July 1, 2015 - link

    Say what you will about the performance, the HBM setup is pretty cool. Hopefully help to bring down card sizes.

Log in

Don't have an account? Sign up now