VR World

NVIDIA Unveils Pascal GPU: 16GB of memory, 1TB/s Bandwidth

NVIDIA Corporation Logo

At the Japanese edition of NVIDIA GTC (GPU Technology Conference), NVIDIA finally revealed details behind its 2016 graphics architecture, codenamed Pascal. The architecture was launched at the main GTC event, which took place in San Jose on March 17th, 2015 (watch Jen-Hsun Huang’s GTC keynote here). GTC Japan was hosted by Marc Hamilton.

As always, the Pascal GPU will be manufactured in Taiwan Semiconductor Manufacturing Company (TSMC), using the brand new 16nm FinFET process. This process is much more than a simple number, since it marks the shift from planar, 2D transistors to the FinFET i.e. 3D transistors. This shift required that the engineers make lot of changes in the thought process, and should result in significant power savings.

NVIDIA Pascal proof-of-concept Engineering Board with 4GB HBM memory. Retail products will carry 16GB of HBM2 memory.

NVIDIA Pascal proof-of-concept Engineering Board with 4GB HBM memory. Retail products will carry 16GB of HBM2 memory.

But that is just the beginning, as Pascal will bring support for up to 32GB of HBM2 memory. However, the actual products based on Pascal will launch with 16GB HBM2 memory, and more memory will depend solely on memory vendors such as SK.Hynix and Samsung. What is changing the most is bandwidth. Both the Kepler-based Tesla (K40) and Maxwell-based M4/M40 featured 12GB of GDDR5 and achieved up to 288GB/s of memory bandwidth. Those 16GB HBM SDRAM (packed in four 4GB HBM2 chips) will bring 1TB/s in bandwidth, while internally the GPU surpasses the 2TB/s barrier.

NVIDIA’s Marc Hamilton said: “Using 3D memory, not only the memory capacity will go up, the memory bandwidth will go up significantly. With a much faster GPU, and higher memory bandwidth, the existing interconnects in the server are just plain outdated. So, we had to develop our own interconnect called NVLink, five times faster than existing technology.”

Pascal will also be available in multi-GPU packaging, replacing the Tesla K80 (NVIDIA skipped Maxwell-gen dual-GPU Tesla). Combined figures are very interesting to compare – 24GB GDDR5 and 480GB/s bandwidth should be replaced with 32GB HBM2 and 2TB/s bandwidth, mutually connected through NVLink rather than PCIe. The NVLink will enable up to 80GB/s, which should replace PLX PCIe Gen3 bridge chips that can only support 16GB/s (8GB/s per GPU). This part should be ‘warm up’ for 2018 and the Volta architecture.

Pascal GPU architecture highlights: Higher performance through 16nm process and GPU Boost, stacked 2.5D / 3D memory, NVLink and mixed-mode calculation.

Pascal GPU architecture highlights: Higher performance through 16nm process and GPU Boost, stacked 2.5D / 3D memory, NVLink and mixed-mode calculation.

Unfortunately, the company did not disclose how much would ECC (Error Correcting Code) reduce the memory performance and overhead, but that is something all HBM-powered products will have to deal with. In any case, the company is gearing for a battle with Intel Xeon Phi, which in its recent incarnation is becoming quite the competitor. Still, Pascal is expected to deliver double-digit single-precision TFLOPS performance, and a lot of focus will be placed on so-called mixed-mode precision (INT8, FP16 and FP32).

Pascal is expected to hit the market during the first half of 2016.

  • Xabi Granja

    Unless I’m dreaming, we’ve known this information for months… how is this news?

    • Broc kelley

      well I have a gxt780 and just recently considered upgrading as I’d like a second PC for streaming. figured why not just build a new gaming pc while I’m at it, this article helped.

      if you have more info than was shown here please link it, as I’m just googling here.

      • Zon

        There isn’t much more information about Pascal than there is in this article. GPU’s would be at 14nm for the Pascal series if it wasn’t for the demand for smart phones. Smart phone manufactures usually get first dibs now because there is WAY more demand in smartphones than GPU’s. The manufactures that make the 16nm FinFET chips can not keep up with the demand at the time and the graphic card industry is left with the out dated stuff. That’s why NVidia Maxwell and AMD Fury GPU’s are on 20nm FinFET and not 16nm. Now that 14nm FinFET is out, AMD and NVIDIA can use 16nm FinFET.

        • Kenneth Tissen

          What the hell are you talking about?
          All GPU’s up ’till now are made on the 28nm process, FinFET hasn’t been used in GPU’s either.
          The next generation of GPU’s will be using that on either the 14nm or 16nm process

        • Do not agree with you – TSMC is not doing 14nm, TSMC is currently at 28nm and 20nm planar nodes (for leading edge, we won’t discuss 65/90/130nm stuff here), and they’re moving to offer 16nm FinFET. Both Fury and Maxwell were manufactured using 28nm PLANAR (i.e. ‘2D’, conventional) transistors. AMD and NV are yet to launch their FinFET (i.e. ‘2.5D’, ‘3D’) products… mostly due to the fact that their foundries did not offer the process node/technology.

          • Zon

            Sorry memory was a little hazy when I posted it. Thought they were on 20nm and not 28nm. Smartphones and tablets that go through TSMC took a huge chunk of their supply and pretty much didn’t have enough for AMD and NVIDIA. That’s what I remember why AMD and NVIDIA are still on 28nm.

          • While it is easy to say 20, 22, 28, 32nm and so on, the truth is that the processes used by the mobile chip vendors – and high-performance silicon is entirely different. From the outside, it is simple to say 28nm silicon, but the truth is that we have 28nm ULP (Ultra Low Power), 28nm geared for analog chips, 28nm LV, 28nm HP, 28nm UHP etc. GPUs we all know utilize high-power, high-performance processes (HP, UHP), and here’s where TSMC did not offer 20nm HP. With AMD’s decision to switch foundries and skip 20nm – going directly to 14XM (sorta 20nm meets 14nm FinFET) at GlobalFoundries, NV was left with nothing else to do but to jump on the 16nm HP node at TSMC. Take a look at for example, NV chip lineup – Tegra and GeForce/Quadro/Tesla do not use the same manufacturing process.

            Hope this helps.

      • Xabi Granja

        There’s no much more than what you read here, but there’s snippets that indicate movements towards launch. It’ll probably happen by March 2016. You definitely want to wait for Pascal – I’m on a 770 so trust me, I get the impatience. Just HBM2 is a big enough deal to wait for the upgrade, as anything else will quickly be outdated in this new generation of consoles/GPUs with 8GB VRAM. If you go with a 4GB GDDR5… you’ll probably have problems not far down the line.

        Check here for the most up-to-date stuff: http://wccftech.com/nvidia-pascal-volta-gpus-supported-geforce-drivers-adds-support-vulkan-api/

        • Key thing with memory is not capacity, but latency. HBM and HBM2 are moving things in a really good direction, reducing the latency when compared to GDDR5, and offering much higher bandwidth. Unlike system (DDR) memory, which seems to go from bad to worse with every passing generation – clocks are getting higher, but so is the latency. From 200MHz DDR clock with CAS1.5 (DDR1-400) to DDR4-2133 with CAS15. Sigh.

          • Xabi Granja

            You’re right about latency, but don’t dismiss simple capacity so fast. If you’re playing a game that requires, say, 5/6gb of VRAM (say, AC Unity which jumps upward of 4GB) and you only have 2, performance will quickly drop to the floor because there’s not enough memory to hold all the textures, etc. Latency is key, but if you don’t even have the capacity, latency doesn’t matter at all.

            Thankfully, HBM solves both problems 🙂

          • Nobody is dismissing capacity. Capacity is a paramount, but when it comes at a cost of high latency and significant power consumption… sometimes speed is everything. Then again, the requirements of usage models differs greatly. For example, AMD is not launching its professional parts based on Fiji GPU with the fear of 4GB being too low, yet initial results speak of at least one client that will use consumer cards in a professional / commercial environment due to the sheer bandwidth offered. Thus, it all depends from what view are we looking into GPU and CPU hardware. HBM1 is good, but ideally, we need 8-16GB HBM2 memory on cards. Will we get it? Sure thing. But I would rather have all the cores working at 90% and 4GB of memory, rather than having 8GB of memory and unoptimized code tasking only 75-85% of the GPU.

        • Godsbane777

          I will have trouble with only 6GB of VRAM soon. Looking forward to Pascal having 16GB and 32GB. I wonder how well they will scale in SLI. Three and four-way SLI with good scaling will open the door for incredible resolutions with all that VRAM available.

        • Duxa

          You guys are crazy, Im on GTX 470 and Fallout 4 is the first game I cant run on Ultra. whats with the impatience to upgrade? Ill be upgrading to Pascal but thats because my hardware is 6 years old. If I had GTX 670 or newer I would wait until Volta.

          • Xabi Granja

            That simply can’t be true. I jumped from a 470 to a 770, so that’s 3 generations. My 470 couldn’t sustain 1080p on high settings in AC4 Black Flag or Bioshock Infinite, much less something like AC Unity or the new games we’re getting now. It’s simply not possible, so I don’t believe your statement as I had the same GPU as you and Ultra on 1080p was untenable, period.

            The 770 could get me by for another year easily, but VRAM has now become a problem with the newer consoles and if you want full detail on textures, you need 4+GB of VRAM, thus my 770 comes up short.
            Honestly though, I might switch to the x60 range because I’d spend the same money in 6 years buying x70 range every 3 years vs x60 range every architecture change aka 2 years (300+300 VS 200+200+200) and it seems like 1080p doesn’t usually need much more than x60 level (unless it has less than 4GB VRAM, in which case I’m going with x70)

          • Duxa

            To be fair I havent played either of those games on my GTX 470. Nonetheless I stand on my point that upgrading from GTX 780 at this point is unnecessary.

          • Xabi Granja

            I would agree on that point. A 780 is enough until Volta, easily. Unless limited by those 3GB of VRAM – which could be a problem depending on the game – there’s no need for now. If you’re on Kepler though… again depending on what you game, Pascal is definitely an upgrade to consider.

          • Crippled by UK State goons

            If you are using a 1080p monitor, there is really no need to upgrade from a GTX780 yet… unless you always use Stereoscopic 3D or Multi-monitor… Few games use more juice than a 780 can handle at Ultra settings (I could mention a couple; but why).
            I do have around 400 games on Steam/Origin/UPlay, and am actually looking forward to the day when my 780 is for dedicated PhysX.

          • Jesse Colts Lafantaisie

            Some people have 144hz monitors or run their 144hz monitor on 120hz lightboost that means they need 120 fps or more at all times an if on 144hz they need 144 fps or more. Also these monitors are 1080p so yes if you are using a 1080p monitor there is reason to upgrade, because I highly doubt a gtx780 can max out every game and get 144 fps constant. Not everyone is on 60hz and only needs 60 fps.

          • Crippled by UK State goons

            Stereoscopic 3D is 120Hz, not 60.

          • And 144Hz. 🙂

          • Crippled by UK State goons

            Well, I am still waiting for the main runner to emerge in the VR world, for a change to be really worthwhile.

          • It will still take some time to get a major player, but if I would have to pick, Oculus has serious muscle behind them, stuff Valve is doing is also impressive (more on that later), and there are several pretty heavy market cap players which might push their vision through. Trouble is, you need more compute power, dual GPUs and all that jazz, and the industry hasn’t adjusted to the realities of new markets. If a $1000 cellphone, which an high-end iPhone or Galaxy S6 are – are being sold for 20 bucks a month for two year contract, why a $2000 PC could not come for $50 on your telco bill? Pricing issue, and the aura of high-end graphics being ‘expensive’ is the key blockage of VR becoming mainstream. We need VR headsets in the living room for $10/mth (which a $299 set from Oculus or Valve certainly are), PC for $50/mth and the adoption is ahoy.

          • lance vance

            Exactly this

  • Pingback: NVIDIA unveiled a graphics card with 16GB of RAM and now I can't look my own PC in the eye - The Core Gaming NetworkThe Core Gaming Network()

  • Pingback: Intel Moves to Attack NVIDIA Quadro, AMD FirePro Market | VR World()

  • charlieboyk

    using the brand new 16nm FinFET process…….. meaning 1 in 100 dies will be working correctly, nVidia will be able to ship consumer quantities in 2025.

    • Wishful thinking. Competition is good, and we will see AMD at 14nm vs. NV on 16nm. Things are bound to get interesting, especially when Intel joins the frame.

      • GregoryGr

        Intel will never join the dedicated gpu market the way things stand now. They can only produce stuff that would be on par with a decade old NVIDIA/ATI cards. They tried in 2010 and the results were so underwhelming for them they completely dropped the project.

        • That used to be true, but the realities today are different. Ready my story from yesterday: http://vrworld.com/2015/11/17/intel-moves-to-attack-nvidia-quadro-amd-firepro-market/
          They are coming, and they’re coming sooner than you think. A single Xeon E3 (read: mainstream Skylake or Broadwell based parts) paired with an Xeon Phi.

          • GregoryGr

            If they somehow manage to get on rails and follow Nvidia and AMD that would be a good day for the consumer. AMD will have more of a reason to make something good for a change and Nvidia will think twice before pricing. I’m game.

        • Akemi

          but Intel could grab the low end gpu market from both AMD and Nvidia. nowadays almost no one uses low end GPUs of Nvidia/AMD unless it comes with a laptop..

    • Intel and Samsung are already at 14nm.

  • Pingback: A peek at Nvidia’s 2016 GPUs | writingforinfoprofessions()

  • RedSocks157

    If this thing is real…there goes AMD’s play for the high-end GPU market.

    • Mai Johnson

      This is a professional card at best. The cost will be astronomical. 16GB is overkill even at 4k.

      • Akemi

        if highest is 32GB, the lowest may be 2GB. means future games will be more quality.

        • Don’t think 2GB is possible. HBM1 memory is 1GB chips, HBM2 brings 4GB chips. 2GB will probably be possible through GDDR5 and GDDR5X memory, but that is very doubtful. 4GB will be the minimum going forward. If not for performance, for ‘sticker effect’ on boxes.

          • Akemi

            yeah my point was future games will get more memory than now.

        • Duxa

          Games will still be bottlenecked by consoles since thats where most of them sell.

          • Akemi

            texture quality isn’t an issue as developers can always release a higher quality textures specially for PC like Witcher 3, GTA ,Saints Row Three.

          • Duxa

            Yup; however in most cases thats unnecessary work that most developers are not going to do since it does not impact sales in any meaningful way. While yes original textures are higher rez then scaled down during optimization you cant just release the high rez ones, they have to be play tested etc, a lot of extra work for little to no profit. So while some devs do it and we praise them for it. Most wont.

            Mods will however as they always have 🙂

            If you look at previous generation which features 512MB of VRAM (360 and PS3) that held PC gaming back significantly, most games recommended spec was under 1GB of VRAM. Once we got XBONE and PS4 we suddenly jumped to 2 to 4GB recommended. I dont expect a jump beyond 4GB until next gen of consoles 5+ years from now.

            Will we benefit from extra RAM? sure, mods and some devs will be awesome and release high rez packs… but for most part 4GB should be plenty for years to come.

          • Akemi

            you have a point but we can’t predict the future based on the past because everything is depend on the manufacturing cost. with HBM2 you can go up so easily and it’s cheap to produce too.

      • $649 price point should fit nicely. Don’t forget that with HBM/HBM2 memory, adding and deducting memory channels is much simpler affair than it was with the conventional ‘drive every channel’ and ‘transistors for memory controllers’. I remember the 512-bit ringbus architecture on Radeon 2900 and the complex 384-bit on the first GeForce. It’s not known, but today the memory controller is simpler and performs much more than in its initial iterations… AMD and NV both did great work in reducing complexity and increasing performance.

    • Just posted what AMD is planning for 2016. Check the AMD Greenland GPU here – http://vrworld.com/2015/11/19/14nm-amd-greenland-tapes-out-attack-on-nvidia-pascal-intel-xeon-phi/

  • Brad

    This is kinda making me regret getting my 980 Ti…

    • Godsbane777

      Never regret your 980 Ti, only love. Simply regret not buying that extended warranty.

    • You can sell it on ebay and take a 50% loss. It’s not even out so it’s worth having until it’s released.

    • Akemi

      sell it before a lot of people know about this news.

    • GregoryGr

      what were you thinking exactly when you bought a card based on a product line that was released a year ago? You know they make cards every year right?

      • Brad

        Well when I got it 6 months ago, I was sure that consumer VR would be any day. I was only off by 8-12 months. But I can’t go back to my old 750 Ti. Star Citizen 2.0 won’t accept that.

        • Me Myselfandi

          Yeah if I had my old 280X laying around I’d sell my 980TI now for the newest card, either AMD or Nvidia, but I don’t, I gave it to a friend lol. So I can’t sell my card as I’d have no graphics card.

        • GregoryGr

          if you bought it 6 months ago it was worth it at that time. If you bought it yesterday that would be a fucking stupid move.

    • Why would you regret it? You will have amazing performance, and then a next generation product will come out, which will bring even more amazing performance. When you feel the time is right, you will upgrade. In meanwhile, start the game / application you love the most – and enjoy.

    • Duxa

      You cant chanse tech. If you would have waited and got Pascal then there would be news about Volta that would make you say the same thing, then if you waited for Volta whatever comes after would put you in same predicament. Just enjoy your 980Ti, you are set for at least 4-5 years with it.

    • Crippled by UK State goons

      Mmmmm super dedicated PhysX card.

  • Godsbane777

    Dying for the 1080 Ti in SLI. Bring on 5K gaming!

    • Blue Gum

      Gaming at 4K and beyond will only really start to be standard once devs make the switch to DX12/Vulkan.

      • Godsbane777

        This time next year is the estimate of DX12 implementation in modern games. These Pascal cards look incredible. I believe by then I won’t be in the minority gaming in Ultra HD. It was always said Maxwell, despite being a great architecture, is really just a small step forward compared to what Pascal is going to bring. Maybe that is going to be even more true than I thought.

    • I’m more thinking about the 1080 GTX for laptops.

      • Godsbane777

        Looks like Pascal is really going to deliver. Potentially five times more VRAM in 3D stacks.. If the processing cores improve by that same margin, even remotely, then 5K gaming is going to be a joke.

        • 5K is not going to happen. We’re in the transitioning period. In the display business, everyone is talking about 4K and 8K. 4K is being considered to what 720p was compared to 1080p. 8K is the target resolution for TV manufacturers, and the prices are already in the $25K range on the high end (105″ model). We’re few years away from seeing 8K for $2K, but that time will come. Unless… we really get into the VR and AR worlds, with AR being much more demanding than VR. Pascal and Greenland GPUs are both architectures for the future, and I for one cannot wait to see EVE: Valkyrie.

        • 5K is not going to happen in the way most people think. We’re in the transitioning period. In the display business, everyone is talking about 4K and 8K. 4K is being considered to what 720p was compared to 1080p. 8K is the target resolution for TV manufacturers, and the prices are already in the $25K range on the high end (105″ model). We’re few years away from seeing 8K for $2K, but that time will come. Unless… we really get into the VR and AR worlds, with AR being much more demanding than VR. Pascal and Greenland GPUs are both architectures for the future, and I for one cannot wait to see EVE: Valkyrie.

          • Godsbane777

            I think 6k will be what comes next year. These Pascal cards look to be a tremendous step forward. 5k will be and is already present, but I can use DSR to run 5k no problem with twin 980 Ti’s. Pascal’s 980 Ti and Titan X equivalent will be 6k capable in SLI. I really believe that.

          • lance vance

            I’ve seen a couple of 8k screens and they’re crap. The real new benchmark for impressive displays of the future is higher and higher ppi’s that’s what I’m looking forward to.

  • Pingback: 最快 2016 年第一季亮相,NVIDIA 計畫為 Pascal 加入 NVLink 與 HBM 2 - BenchLife()

  • Akemi

    my Skylake processor is already with 14nm 😀

    • AJ Lee

      Mine too, 6700K :3

  • GregoryGr

    I’m not sure i interpret this successfully, but PCI 3.0 is not going anywhere in the close future right? because if so you can forget my money Nvidia. I’m not going to buy a new socket and CPU.

    • NVLink is primarily considered as an enterprise-class solution. It will trickle down to desktops as well, just give it time. 🙂 PCIe 3.0 is kinda outdated with 16GB/s bandwidth, when the GPUs are achieving 512GB/s, and soon even 1-2TB/s…

      • GregoryGr

        That means what?

        • First and foremost, it is PCIe or PCI Express. PCI already had a version 3.0, and no, there were no graphics cards for it, as they used AGP (we’re talking about 2004 here). PCIe standard is expanding, and it will see a backwards compatible 4.0 – meaning your old hardware will work on new spec boards. 16GB/s is slowly getting outdated, and it will be replaced with 32GB/s bidirectional bandwidth. More information can be found on https://pcisig.com/

          • GregoryGr

            So it will be up to the card to not be powerful enough in order to not be bottlenecked by the current hardware? yeah i can’t see that happening unless i buy a $600-700 price range card which i have no plans for.

          • In that case, Pascal, and Greenland – are not products for you. Just like probably, Haswell-E, Broadwell-E are not the products you will base your system on?

          • GregoryGr

            my system is already based on Haswell. I have a 1150 motherboard an i5 4670k cpu and 16gb of 1600mhz DDR3 RAM. This pc is not going anywhere for the next 5 years at a bare minimum and if Nvidia knows whats good for them, they will allow me to buy their new products without having to replace any other part. I’m sure the majority is not waiting for pascal to hatch out to build their systems. Most of them just want to fit in a new gpu.

          • I am quite certain you will be able to buy consumer parts compatible with your platform going into the future. NVLink and HyperTransport are geared towards server implementation. It will come to desktop, but that does not mean PCIe will not continue in foreseeable future. How much the CPU will hold back the new GPU is another thing.

          • GregoryGr

            Depends on the gpu. I’m guessing something as powerful as a 980ti will not be held back by the likes of my cpu. I don’t think that a 300-400 budget will even get me that high.

          • Crippled by UK State goons

            I am sure a GPU mounted `CPU function chip` to help an older on-board CPU chip to work 100% with any new tech can be developed.

          • Crippled by UK State goons

            Same here… and I have a couple of them.

  • Pingback: NVIDIA revela más detalles de Pascal en la GTC Japón, hasta 1TB/s - Tecnius()

  • ded2me

    Just bought a 970. I wont upgrade for at least 3 years. And the Game makers know that.

    • Duxa

      You are set until next generation of consoles in about 5 years.

      • Crippled by UK State goons

        Sure; at 1080p. And not steroscopic 3D, or multi-monitor.