1. TechSpot is dedicated to computer enthusiasts and power users. Ask a question and give support. Join the community here.
    TechSpot is dedicated to computer enthusiasts and power users.
    Ask a question and give support.
    Join the community here, it only takes a minute.
    Dismiss Notice

Final specifications for the AMD Radeon Pro Duo announced

By Scorpus · 35 replies
Apr 26, 2016
Post New Reply
  1. AMD already announced the Radeon Pro Duo earlier this year, but today we're finally getting all the juicy details about the dual-GPU graphics card the company is aiming at virtual reality creators and consumers.

    As expected, the Radeon Pro Duo features two fully-unlocked Fiji GPUs, each with 4096 stream processors, 256 texture units, and 64 ROPs. Along with a core clock speed of 1,000 MHz, the performance of a single core should be around the same ballpark of the Radeon R9 Fury X, which is clocked slightly higher at 1,050 MHz.

    The card is loaded with 4 GB of HBM per core, providing a ton of bandwidth but perhaps not quite enough memory for some of the highest end games the Pro Duo might be expected to run. Interestingly, the Radeon Pro Duo has a rated TDP of just 350W, which is the same as two R9 Nanos, despite having three PCIe power connectors on the card itself.

    The Pro Duo's cooling solution is a closed loop liquid cooler provided by Cooler Master, which covers both GPU dies as well as other supporting chips found around the board. The external radiator is 63mm in depth and includes a single fan.

    The pricing for the Radeon Pro Duo hasn't changed since it was announced: you'll need to fork out $1,500 for this card, which is around $500 more than what it would cost for two single-GPU R9 Nanos. It's unlikely that this card will sell in significant quantities, although AMD can now claim that they have the fastest graphics card on the market.

    Permalink to story.

  2. Adhmuz

    Adhmuz TechSpot Paladin Posts: 1,769   +598

    "AMD can now claim that they have the fastest graphics card on the market."

    This will be short lived, Pascal is out in June and anyone who's intelligent will ignore this monstrosity until then, take the win while you can AMD.
  3. VitalyT

    VitalyT Russ-Puss Posts: 3,456   +1,734

    I believe that's the entire point of this otherwise a useless exercise of power vs efficiency.
  4. hahahanoobs

    hahahanoobs TS Evangelist Posts: 1,915   +595

    HBM + Water = "Up to 1000MHz"
    Either Fiji yields are amazing, or next to no one is buying them.
    I predict a price drop very soon.
  5. Nitrotoxin

    Nitrotoxin TS Addict Posts: 109   +62

    but with 4GB of RAM it's a pile of crap...
  6. Evernessince

    Evernessince TS Evangelist Posts: 1,786   +1,007

    Pascal is purported to release in June although we don't know for sure yet. What we do know it that Nvidia and AMD are not releasing the big boys right out the gate so this card will remain on the top spot for some time.

    Seeing as Nvidia's top end GP100 die size is so big, even when Nvidia does release their top dog it will be in very low yields for a long time. That combined with a 50w increase in TDP means that the card is mostly going to be seeing gains from transistor count. It's going to run a good bit hotter than Maxwell.
    mlscrow likes this.
  7. Frixx

    Frixx TS Rookie

    Too little too late for too much. Next.
    Nitrotoxin likes this.
  8. dividebyzero

    dividebyzero trainee n00b Posts: 4,891   +1,262

    As a single card solution it will, but unless Polaris 10 can't approach AMD's Nano in performance and/or costs $750+ each at launch the Pro Duo will remain a prosumer part attractive to VR developers for the 24/7 software support. The title of fastest card is academic interest for the most part. AMD lack of plugging the card for consumers - no gaming based reviews by mainstream tech sites, limited availability, and very heavy emphasis on pro support (AMD is offering FirePro drivers for the card) puts it fair and square in a niche market - I'm not sure how many gamers would sign up for it with 14nm/16nm cards due to be unveiled in a months time.. The no review and pro driver upselling even take it out of Titan/Titan Z territory and plant it in workstation card territory which tends to dilute the PR.
    You are assuming that the consumer GeForce card is based on GP100. Please note that GP100 is not architecturally set up for gaming.
    1. Each die devotes space for four NVLink interfaces. A maximum of one might be seen as advantageous if Nvidia does away with the Avago PLX bridge chip for dual cards - but not four.
    2. GP100 has 1,920 dedicated FP64 units whose power budget would equal the 3,840 ALUs. More likely would be Nvidia reworking the design to cull FP64 and increase FP32 cores - the oft mentioned GP102 whose core count seems to bounce around depending on the importance of FP64 for a consumer card on the person making the estimates seems a more likely product. Culling FP64 reduces power budget, and omitting NVLink which can't be natively used on consumer PCI-E systems reduces die space and pin-out.
    3. GP100 has a cache structure that mirrors the GK 210 - a GPU that only ever saw the light of day as a powerful compute only card.
    Burty117, madboyv1 and mlscrow like this.
  9. mlscrow

    mlscrow TS Enthusiast Posts: 25   +20

    First of all, AMD has had the fastest graphics card on the market for almost 2 years now. Where have you been? The HD7990 still destroys a Titan or 980Ti. With this, AMD will have the fastest 2 graphics cards on the market.

    I doubt this will be short lived either. NVidia's Pascal is coming out, but not before Polaris from AMD. And if and when NVidia tries to launch their next Titan, AMD will already have Vega to trump. AMD has been planning this comeback for years and now they're executing and executing well at that. Sorry NVidia fans. It's about time!
    Last edited: Apr 27, 2016
  10. mlscrow

    mlscrow TS Enthusiast Posts: 25   +20

    Wow, these forums are terrible. You people don't really know much about what's going on. This is not a "useless exercise of power vs efficiency". AMD has already had the fastest GPU on the market for over 2 years now. They created a great GPU with the Fiji chip, which, in terms of price/performance ratios, is a better buy that the GTX980Ti, especially in DX12 titles (which is all titles being made today and into the future), where it outperforms the 980Ti and in some cases the Titan as well.

    The reason for this card, specifically is VR. In order to obtain the specifications required for VR not to induce motion sickness, you need at least 90fps. to get 90fps consistently, with high resolution and graphic, you absolutely need two Fiji chips, one dedicated to each eye, as no single chip GPU today is capable of achieving that level of performance. This is where the ProDuo comes in. Considering this is the year that VR breaks out with Oculus Rift, Sony's, Microsoft's, Samsung's, and many others VR headsets being released, along with the fact that AMD currently holds 83% of the VR market and is one of VR's biggest supporters and drivers, the release of the ProDuo is nothing less than fitting.

    Also consider that an NVidia Titan costs $1500, then compare that to a Pro Duo which costs the same, and the fact that the Pro Duo absolutely obliterates a Titan in terms of performance, and AMD has one incredible card on their hands. I am looking forward to purchasing a VR set this year and honestly, I may very well just pick up this card for it. NVidia will release the 980 successor in the 2nd half of this year, not to include the successor to the Titan, (GP104 will be release, but not GP100), meanwhile, AMD will be releasing the successors to the mainstream and low power markets with Polaris 10 and 11 respectively. AMD will not release a successor to the Fury (as they just released it along with the Pro Duo and want it give it some time under the limelight, before replacing it with Vega, which will most likely be next year.

    Even with these cards being released, achieving 4k@90fps on a VR headset will not be possible without multi GPU's. It won't be until next year or the year after that we achieve that type of performance out of a single gpu die.

    Again, considering that AMD is leading the way with VR technology, they themselves are working on and have as a goal, a VR headset that powers two 16K monitors and maintains 90-120fps. Now THAT is a respectable bar for a company to be reaching for. AMD is doing well for once, people should appreciate this.
  11. mlscrow

    mlscrow TS Enthusiast Posts: 25   +20

    You obviously don't know anything about how GPU's work. a 980Ti needs 6GB of VRAM, because it uses GDDR5. A Fury only needs 4GB, because it uses HBM, which is 9x faster than GDDR5, thus allowing a Fury with "only" 4GB of RAM to outperform a 980Ti with 6GB of RAM once you get up to the high resolutions like 4K. As HBM manages memory in a different and more efficient way, not as much of it is required to maintain performance. The Pro Duo is an incredible card, especially when compared to the equally expensive Titan, which is slow in comparison.
  12. mlscrow

    mlscrow TS Enthusiast Posts: 25   +20

    Nice, someone here actually knows what they're talking about. Thank you. These people need to be educated.
    Evernessince likes this.
  13. Nitrotoxin

    Nitrotoxin TS Addict Posts: 109   +62

    Umm no.... I can dump 6 gallons of water into a 4 gallon bucket and its going to overflow no matter what...sorry but your logic is flawed... It doesn't matter how fast the RAM is when the game is programmed to grab 6GB's of high res textures and there is only 4GB of RAM, it's going to stutter...sorry...
  14. mlscrow

    mlscrow TS Enthusiast Posts: 25   +20

    Wrong. Please do some more research, get the facts down, and then come back with an argument for your position, which you won't, because you'll learn you were wrong to begin with. A 6" wide hose, shooting out 20psi of water, performs less than a 4" hose shooting out 200psi of water. Make sense now?
  15. mlscrow

    mlscrow TS Enthusiast Posts: 25   +20

    Just to prove my point. Here is a benchmark of Shadows of Mordor, one of the most VRAM intensive games in existence today. As you can see here, at 4K resolution, the Fury X with its 4GB of HBM performs the same as a 980Ti and Titan with their 6GB and 12GB of GDDR5 respectively. At this level of resolution and image quality, all 6GB are used on the 980Ti, yet...the numbers speak for themselves.

    Last edited: Apr 27, 2016
  16. mlscrow

    mlscrow TS Enthusiast Posts: 25   +20

  17. Theinsanegamer

    Theinsanegamer TS Evangelist Posts: 634   +605

    *facepalm* That graph does not show any VRAM usage numbers, only framerate. If what you are saying was somehow true (spoiler alert, it isnt), then why doesnt the 390, with 8GB of GDDR5, beat the fury x and 980ti? The only thing that graph shows is that the fury x and 980ti are tied for performance in that game.

    HBM has more bandwidth, but that does not reduce the amount of VRAM used for holding textures, assets, ece. Also, many games nowadays will use additional VRAM to cache in textures/assets that will be needed in the future, as there is room available. GTA V and battlefield both do this.

    Come back when you learn how game programming and VRAM work.
  18. mlscrow

    mlscrow TS Enthusiast Posts: 25   +20

    Wow, the sheer amount of stupidity in this forum is ridiculous. The graph shows exactly what I said it shows, PERFORMANCE. If 4GB wasn't enough, you would see lower fps than the two NVidia cards, that generally outperform the Fury in DX11 titles. This one in particular uses all 6GB of DDR5 that are on the 980Ti, YET, the Fury X outperforms it. This shows how 6GB isn't necessary when you have HBM, you can get away with 4GB and still perform just as well, if not better than 6GB of GDDR5. The VRAM requirement for the game is 6GB if you're running 4K at Ultra settings, which is what I showed in the graph, genius. If you really want to understand, here's a video, showing the exact same game, on Ultra settings, but only at 1080P, and it still uses over 5GB of VRAM. I know very well how game programming and VRAM work. I'm surprised you don't, considering you speak as if you are a programmer.

    You don't get it and I really don't care that you don't. Go on believing whatever you want all while facepalming as if you know something, it'll just make you look even dumber.
    Last edited: Apr 27, 2016
    Evernessince likes this.
  19. mlscrow

    mlscrow TS Enthusiast Posts: 25   +20

    And umm...If a 980Ti performs worse than the FuryX with 50% more RAM, and the Titan with 200% more RAM, then why on earth would you think that 8GB would somehow, miraculously be better or interpret that from what I wrote? Here's a spoiler alert for you..."The 390 DOESN'T have HBM, genius", which I thought a "TS Enthusiast - Programmer - VRAM expert" would know, but apparently, my instincts were correct and you have no idea what you're talking about.
  20. alexnode

    alexnode TS Enthusiast Posts: 35   +12

    Vram is used to store textures mostly and in 4K textures can be a few megabytes each, you can fill 4GB extremely quickly. Even if the memory is fast you have to read and unload textures really quickly. These textures have to pass through your system somehow since they can't be all preloaded on the vram. I guess most games for consoles are really last generation but if you are in for 4K gaming in the maximum quality or thinking of catching up with the games to be released in the next couple of years this 4GB of ram is not smart for the price. It could work well for first generation VR but for other uses it could be silly.
    Anyhow I never got this idea of two cards in one If they could share memory somehow having 8gb shared that could be fun.
  21. mlscrow

    mlscrow TS Enthusiast Posts: 25   +20

    I generally agree, however, at this moment in time, 4GB is just about on the cusp of being not enough, but we're not quite there yet. Granted, 4K hasn't become mainstream yet, but even titles that are VRAM hungry like Shadows of Mordor, when run at 4K with the Ultra texture pack, which has a requirement of 6GB of VRAM, which pretty much excludes most cards outside of a 980Ti and Titan, seem to process just fine under the Fury's 4GB of HBM. The video I posted earlier shows that even at 1080P with Ultra Textures, over 5GB is being used. However, just because a certain piece of software makes it seem like x amount of VRAM is being used, doesn't necessarily mean it's true. 6GB can be "called upon" to be used, but that doesn't actually reflect specifically how much of that memory is actually in use.

    From the research I've done, it seems that it's still a debatable subject on whether or not 4GB of VRAM is enough for gaming today, including 4K resolution. The answer that the results lean to is "yes, it's enough", but it's still rather inconclusive. Some titles show AMD at a disadvantage, however, those titles are also Gameworks titles that specifically cater to NVidia and purportedly hinder AMD performance, so the conclusions from those benchmarks, although they can be considered, can't really be the basis for a solid conclusion. The rest of the titles outside of Gameworks, seem to run just fine on the 4GB of HBM that the Fury is equipped with, even though the NVidia counterparts push beyond 5GB. In some of these benchmarks, AMD starts to outperform the 980Ti the closer you get toward and including 4K resolutions, which would suggest that the 4GB of HBM is indeed enough.

    What I've come to conclude from this research is that generally yes, 4GB of HBM is enough for today's gaming standards, up to an including 4K, and for at least the first generation of VR, which most likely won't be hitting 4K on each eye. However, I do think that this should be the last generation for 4GB on a flagship/high-end type card. With the release of HBM2, we should be seeing 8GB in it's place, which will most definitely be enough for 4K.

    To think that AMD considers the ProDuo not only to be a gaming card, but also a professional card, while maintaining 4GB per die, which is relatively low, compared to the FirePro they recently announced which sports 32GB (most of any card to date), is indeed a bit odd. Either AMD has made a mistake (unlikely as the current AMD team is pretty solid in terms of engineering), or they've calculated everything extremely precisely and determined that 4GB is indeed enough for 4K resolutions (which is why they boasted the Fury as a 4K gpu), as well as VR development, while keeping costs as low as possible. I tend to believe the latter is true, but I cannot say with certainty.

    Also, keep in mind, this is 4GB per Fiji die, so in total you will have 8GB. Is it shared? No, but at least each gpu get it's own 4GB, and each one of your eyes get it's own Fury. I only defend this card, because it's actually really good.
    Last edited: Apr 27, 2016
  22. mlscrow

    mlscrow TS Enthusiast Posts: 25   +20

    That is the true purpose behind this ProDuo. To have a dedicated GPU for each of your eyes in this first gen of VR hardware. They designed it specifically for that purpose. For approximately the next year, this will be the best card for VR purposes one can get. Outside of VR, having two GPU's process a game, nets huge performance gains. To have them engineered so that both are on the same card also has some advantages in reducing micro-stutter, lowering power usage, heat, space requirements, motherboard requirements, cooling requirements, etc.

    Also, AMD wanted to build its own Gaming system, which also runs VR better than anything else, it is called Project Quantum. They designed one of the most amazing looking PC enclosures I've ever seen, but since they didn't have their own hardware worthy of such an enclosure available, they put the project on hold until they released both this Pro Duo, which they want to use in the Quantum as well as the next Zen FX CPU. That will be one amazing little gaming rig.

  23. dividebyzero

    dividebyzero trainee n00b Posts: 4,891   +1,262

    Thanks to the vagaries of Crossfire scaling, dual cards of either vendor come attached with a heavy caveat. As for the HD7990 it wouldn't come close to a GTX 980 Ti - especially when the newer and more powerful R9 295X2 barely shades the single GPU card:

    Unless you have hard evidence to these" facts" I would suggest you tone down the rhetoric unless this is a drive-by posting blitz and you don't plan on being around when these parts actually arrive. GP104 and Polaris are due to drop around the same time. AMD's Roy Taylor is already playing down AMD's all-out-performance in favour of targeting value for money - this should be a pretty obvious indicator.
    AMD’s Polaris will be a mainstream GPU, not high-end
    You might also note that AMD's Vega has already slipped one quarter on the roadmap. You should also note that the chip's size allied with Globalfoundries rather fanciful predictions (and outright lies) for any of its previous process nodes shouldn't guarantee too much - either time wise or the assumption of volume fully enabled dies.
    Technically, AMD have been planning a comeback since the Evergreen series. The R600 debacle and the hastily reworked R700 bought time to lay the foundation of a more competent graphics strategy - better mGPU scaling, Eyefinity etc)
    Firstly, not all games will be DX12 going forward. Unless they have incentive, developers aren't going to invest more time and resources (read money) into coding for DX12 unless they have incentive to do so. Game devs and studios can barely - on occasion - get a AAA title out the door without it requiring patches and fixes and launch target slippage as it is. Good luck shoehorning architecture/GPU specific coding requirements into that morass until game engines are geared primarily for DX12 and parallel coding for other APIs is eliminated.
    I might also add that a large quantity of new games will use the UE4 engine, so Vulkan will gain traction.
    As for the Fury X vs GTX 980 Ti debate. The Fury X shades the reference 980 Ti, but...1. very few reference GTX 980 Ti's are actually sold these days, and 2. GM200 (over)clocks fairly well, which is why you see so many high clocked cards selling at reference prices. As you can see from the chart I posted above, raising the minimum guaranteed boost improves the numbers considerably. Whereas you can buy these cards clocked out of the box, a Fury will require user input and degree of luck/expertise to push a 10% overclock.
    Yes, the Pro Duo is aimed squarely at VR devs, which is why AMD's marketing goes to great lengths to publicize the 24/7 support (devs are seldom tech enthusiasts - they just want everything to work), LiquidVR and FirePro OpenCL drivers etc.
    Nope. Every Titan (GTX Titan, Titan Black, Titan X) has MSRP'd at $999 in reference form with the exception of the EoL'd Titan Z
    You are incorrect. Game drivers will allocate memory usage and prioritize for vRAM depending upon capacity. As you can see here, the larger framebuffer of the 980 Ti means that more of the graphics elements are stored in vRAM, where the smaller framebuffer of the Fury requires elements to swap in and out of system memory. While it appears that the 980 Ti is using more (v)RAM it is because it is less reliant upon system memory.
    That isn't what is actually happening. As resolution/texture computation increases the Fury's greater texture unit (TAU or TMU) count is coming into play. Fiji has 256 TMUs ( 268.8 GTexels/sec), while the 980 Ti has 176 TMUs ( 176 GTexels/sec @ base clock, 189.4 GTex/sec @ average boost). Likewise, at lower resolutions the Fury suffers in comparison to the 980 Ti because the TAU/TMUs aren't the limiting factor, the raster back end is - and Fiji is lacking in ROPs ( 64 compared to GM200's 96 - and each of Fiji's ROPs service a greater number of ALUs (cores)).
    It has nothing to do with memory type or capacity (and please quit the insults and double posting while you are at it). Adding vRAM capacity is only beneficial if the smaller capacity it is being compared to is saturated - I.e. hi res texture packs could choke a 4GB card, but sit comfortably in an 8GB equipped card. Excepting fairly rare instances like this, very high resolutions, or downsampling/full screen AA. It isn't much of an issue. Your assertion that internal bandwidth is a defining factor is just straight out wrong. Very few modern high end GPUs are bandwidth starved.
    The real reason that the (stock) 980 Ti performs worse than the Fury X at 4K
    The 980 Ti has 45% less TMUs
    The 980Ti has 45% less cores
    The reason that the 980 Ti performs better at lower resolutions:
    The 980 Ti has 50% more ROPs
    To that you can add a plethora of other variables including how each architecture handles tessellation, compression/decompression algorithms and how each architecture is impacted by uncompressible content.

    One further note: If you are planning on being condescending and rude, it might pay to actually have the facts on your side. Your knowledge of vRAM/system RAM allocation and GPU architecture isn't very thorough to say the least. What you are presenting is misinformation and in many cases contradictory misinformation. Technical discussions are good, but plastering the thread with vendor specific marketing and assumptions based on little or no factual evidence does nothing to advance the dialogue.
    Last edited: Apr 27, 2016
    cliffordcooley, Adhmuz and Burty117 like this.
  24. Dragonstongue

    Dragonstongue TS Rookie

    Also to add to all of that back and forth mess above.

    AMD has been developing complete systems for years upon years, whereas Nvidia is historically GPU only, and Intel of course is CPU only, so in my mind, AMD has a much better chance of say figuring out how to "stream" the info from the game/app to the memory more efficiently/effectively then might otherwise be normal, example R9-285 or 380 will be comparable memory usage etc to the "higher end" wider bus cards, using 2gb instead of 3gb and 256bit instead of 384bit for the mem bus.

    So yes, if you are hitting out the memory bus more then memory is available of course it will hurt performance big time, but there is also nothing saying 2 things..1 they have figured out how to effectively use much much higher speed memory plus the tweaks gained from years of developing these cpu-gpu systems/architecture to alleviate some of the limitations of lower memory 2 have highly tuned Fury/FuryX to....

    I am sure get the point, are we not learning every day that goes by what is possible might not have been a short time ago, these products are insanely complex, and am quite certain for those who actually design them, have learned WAY more then we give them credit, and to put quite bluntly, if there was EVER a company to do all kinds of amazing things when they are backs to the wall, it IS AMD.

    Fury/FuryX had to be using the memory HBM in this case EXTREMELY effective, which they have, but of course there are always limitations, not loading the rop/tmu certain ways or whatever, to expect everything to be omg amazing all the time is expecting far to much, and AMD with the 285-290-Fury did some amazing things, whereas Nvidia with the 9xx series, the ONLY reason why they appear "so damn fast" they chopped a great deal of "extras" to make them as efficient as possible, which they most certainly did achieve. I despise them as a company for all the years of BS they have done, though the 200-600-700-900 series truly pushed what was possible in certain ways, even if holding back the market and being as greedy as hell in the process, it will bite them in ***, as this is the last year they had cross license with Intel, now it is Intel-AMD much much more aligned.

    Anywho, cannot always KNOW what is fact or fiction unless we are the ones actually designing these things, which obviously we are not, so therefore what we would normally consider impossible, an engineer somewhere is proving it more probable :)
  25. Burty117

    Burty117 TechSpot Chancellor Posts: 3,042   +793

    I'm sort of surprised you even responded DBZ!
    I assumed he was a troll with all his Ramblings and nonsense spewing out across his multiple posts.

    I swear he only pops up when an AMD article gets on techspots main page?

    Oh no wait, just had a look at his profile, he created this new one just today but the way he writes his comments I could be mistaking him for another troll that normally chimes in on anything AMD xD
    dividebyzero and Adhmuz like this.

Similar Topics

Add New Comment

You need to be a member to leave a comment. Join thousands of tech enthusiasts and participate.
TechSpot Account You may also...