Nvidia Titan V brings the power of Volta V100 to desktops

Late yesterday evening, Nvidia announced the Titan V graphics card at the Annual Conference on Neural Information Processing Systems (NIPS). The Titan V puts the same 815 mm² Volta V100 GPU that's powered Nvidia's highest-end Tesla compute accelerators for some time into a desktop-friendly card. Volta V100 is easily the single most powerful GPU ever, so the Titan V is easily the single most powerful desktop graphics card—or compute accelerator—ever. Getting ahold of that unbridled computing power requires a cool $3000.

That distinction between compute and graphics matters again with the Titan V because unlike the Titan X family before it, the Titan V doesn't simply use a bigger version of Nvidia's consumer graphics chips. The V100 GPU on board the Titan V ships with most of its bounty of compute resources intact. The same 80 Volta streaming multiprocessors (SMs) one will find in the Tesla V100 are operational on this card, for a total of 5120 single-precision CUDA cores (or shader processors), 2560 double-precision shader processors, and 640 tensor cores for the acceleration of AI workloads. The Volta V100 GPU has a total of 84 SMs, but a fully-enabled version of that chip has yet to be seen in a shipping product.

A logical representation of the V100 GPU on board the Titan V

This is undoubtedly the most full-bore compute-capable Titan since, well, the Titan, and its primary destination would appear to be the workstations of AI researchers and other compute-hungry pros. Nvidia does have Game Ready drivers for the card, though, so if you're well-off, insane, or probably both, you can use this thing to push pixels.

The big resource change on the Titan V comes in memory capacity and bandwidth. Nvidia has pared back the memory capacity on the Titan V to 12 GB of HBM2 RAM at 1.7 Gbps clocks, compared to the 16 GB on board the Tesla V100 accelerator. In turn, that means the Titan V enjoys 653 GB/s of memory bandwidth across a 3072-bit bus. By comparison, the Tesla V100 can suck down 900 GB/s across its full 4096-bit memory interface. The Titan V uses a 16-phase VRM with integrated power stages fed by an eight-pin and six-pin PCIe connector, and it'll slip into the same 250W board power that accommodates a GTX 1080 Ti or Titan Xp today. Also, it's gold, baby.

Interested buyers can pick up a maximum of two Titan Vs from Nvidia's online store now.

Comments closed
    • tipoo
    • 2 years ago

    Does anyone know for sure yet if CUDA core, not using tensor cores which aren’t applicable to every situation, can use double rate FP16?

    Previous generations just had enough for tolken support of it, but at less than FP32 speed let alone double rate.

    • End User
    • 2 years ago

    Where is the consumer edition? My GTX 1080 is long in the tooth!

      • NTMBK
      • 2 years ago

      I’ll happily, err… dispose of it… for you.

      • Krogoth
      • 2 years ago

      Wait until Q2 2018 and you’ll probably able to get one.

    • davidbowser
    • 2 years ago

    Since several comments mentioned benchmarks, I figured I would link the ACTUAL use case (machine learning) instead of what most folks here are asking for.

    [url<]https://www.xcelerit.com/computing-benchmarks/insights/benchmarks-deep-learning-nvidia-p100-vs-v100-gpu/[/url<] tl/dr Faster/better, but not as much as one would expect based on the specs. It's possible the machine learning software might need to be refactored to take advantage of the V100.

      • MathMan
      • 2 years ago

      That particular example is for RNN networks. It’d like to see the results for CNN networks before jumping to conclusions.

      CNNs are much heavier in pure large matrix multiplications.

        • davidbowser
        • 2 years ago

        Yep. There are other folks that have benchmarked CNN and the results are pretty close to the RNN, in that they show an improvement, but it just isn’t what people were expecting/led to believe based on the specs. Regardless of the expectations, this is the best GPU available for machine learning, and the improvement over P100 is real, so I fully expect people to buy them in truckloads.

        [url<]https://blog.rescale.com/nvidia-tesla-v100-benchmark-results-on-rescale/[/url<] Again, these are still pretty early benchmarks, so I would take them with the requisite salt. I am really starting to wonder if NVidia should take these public benchmarks back and look at the tensorflow and CUDA code to see if there are code improvements to make. EDIT - I found thinkmate.com had the GPU pricing as part of server configs, so it is $5700 for a P100 and $8000 for the V100. So the overall speed improvements would warrant the price.

      • tipoo
      • 2 years ago

      Makes sense. In Nvidias own presentation linked in my other comment, they said with some optimization they got 3x faster NN training. Not 10x from the tensor performance, since tensor math isn’t the *only* part of NN training.

    • psuedonymous
    • 2 years ago

    That’s… not as awful idea as a stopgap gaming card as it might appear.
    In absolute terms of price/performance it’s moronic, of course, but in terms of value over time I’d expect it to hold its value very well. I can’t imagine Nvidia releasing any other V100-based cards anytime soon even CLOSE to this price point, nor can I imagine them flooding the market with these cards (and risk volume sales of Quadros or Teslas). I expect these to remain in high demand for quite some time, so by the time the great-for-gaming-but-no-good-for-compute GV102-based 1180Ti (or whatever) is released, I’d expect you could probably sell one of these for a rather minimal loss second-hand.

      • NarwhaleAu
      • 2 years ago

      What makes it great, is that in five years time we will be able to buy this much graphics horse power for about $300. In only 1 to 2 years we will be close for $800.

    • NoOne ButMe
    • 2 years ago

    Hm. Pref/watt increase is not very much based on the first, and not conclusive stuff here.
    [url<]https://videocardz.com/74382/overclocked-nvidia-titan-v-benchmarks-emerge[/url<] Granted, that also might not have drivers really optimized for it. Plus this does have the overhead of FP64, Tensor cores and more. Although moving to GDD5x or GDDR6 for mainstream instead of HBM2 probably will add more net. So either unimpressive (from pref/watt increase in gaming standpoint) due to it being all based on just being wider and lower clocks. OR It's just has no drivers and that means 12nm or Volta has brought real pref/watt advantage on architecture. I say it brought none real as the pref/wtt increase may be purely on the wider+lower clocked. Hopefully for AMD it's just it being wider. Hopefully for Consumers, who do matter more, it's just a lack of drivers.

      • DavidC1
      • 2 years ago

      Moving to GDDR5X or GDDR6 from HBM2 likely won’t add. I think it might even do the opposite.

      Taking out FP64 and Tensor cores will reduce the die size and cost to a manageable level, but won’t have an impact on performance.

      The performance increase seems to be 30-40% in both games and in compute workloads. That seems interestingly in line with memory bandwidth improvements, 25% on the GV100 and 20% over the Titan XP(34% over 1080 Ti) on gaming.

    • Rakhmaninov3
    • 2 years ago

    I think I’ll make one of these for my wife for Christmas

      • smilingcrow
      • 2 years ago

      If you buy one of these it will make you a wife.

        • NTMBK
        • 2 years ago

        Krieger-san!

    • tipoo
    • 2 years ago

    Jen Hsun introduced this as “only Nvidia would launch the most powerful graphics card to a room full of the poorest people” to a group of PhDs and academics, lol. Kinda funny in his usual in-jokes with the engineers way.

    [url<]https://www.youtube.com/watch?v=-9YhsAaEpmU[/url<]

      • DancinJack
      • 2 years ago

      to be fair, he said “only Nvidia would launch a brand new generation of GPUs to…”

      But yeah. And, of course, while yes the individuals that do the research could be “poor” their institutions backing them are generally quite well funded.

      still funny.

      Jen Hsun 2020

      • Bauxite
      • 2 years ago

      Grant money yo

    • Srsly_Bro
    • 2 years ago

    It isn’t full power. This is a cut-down card. It’s not full gv100.

      • Jeff Kampman
      • 2 years ago

      You’re right, and this is what I get for reading spec tables rather than the surrounding info.

      • NTMBK
      • 2 years ago

      Nope, one of the memory controllers is disabled.

        • the
        • 2 years ago

        Which also means the L2 cache and ROPs are reduced too a they’re tied to memory controller blocks.

    • bfar
    • 2 years ago

    A new desktop GPU architecture at 3k.

    I love it and hate it all at the same time

    • tipoo
    • 2 years ago

    Babys first P100, for those who can’t quite swing 10K for an AI training card to the boss.

    Crazy expensive Geforce, crazy cheap for the tensor performance, this will be big with academics training neural nets against each other.

    • mcarson09
    • 2 years ago

    Not with that crappy stock cooler!

    It should come with a full copper full cover waterblock for that price!

    • Bensam123
    • 2 years ago

    Anyone find any benchmarks on this yet? As with custom solutions like this it seems as though because the demand is very low, supply is very low, price is very high, game developers and even Nvidia themselves wont go out of their way to optimize them for games. When there are issues it’ll take quite awhile to fix.

      • Klimax
      • 2 years ago

      Unverified:
      [url<]http://www.guru3d.com/news-story/nvidia-titan-v-graphics-card-benchmarks.html[/url<]

        • Krogoth
        • 2 years ago

        It looks legit though.

        People who were expecting a massive jump from Pascal to Volta are going to be disappointed. Nvidia is more focused on GPGPU performance this time around and they already used all of the 14nm TSMC magic with Pascal so don’t expect massive bump in clockspeed if any. I suspect Volta will do much better with cryptocurrency mining since Nvidia wants a slice of that lucrative pie.

        I suspect the most of the excitement from customer-tier Volta will likely come in their mid to low-end SKUs a.k.a GV106-GV108 where there’s still room for growth for adding more blocks.

          • tipoo
          • 2 years ago

          I saw some less technical sites run with 110tflops without context, leading to people freaking out over it being 10x faster lol. It would only be for AI training/other tensor math, main cuda core performance is only moderately up.

          • DancinJack
          • 2 years ago

          You say that same line about every NV generation. Who are the people that were expecting it Snoozer, I mean Krogoth?

      • tipoo
      • 2 years ago

      Who is getting it for a modest boost to games though? It’s almost entirely an AI training card with that tensor performance. The CUDA core performance is only moderately up from a Titan xp.

        • smilingcrow
        • 2 years ago

        FP64 is 16 times higher at the base level though.

          • tipoo
          • 2 years ago

          Also that, but again, nigh unheard of in games.

            • smilingcrow
            • 2 years ago

            This clearly isn’t aimed at gaming so irrelevant really.

            • tipoo
            • 2 years ago

            Agreed, but it was the context of OPs post.

            • smilingcrow
            • 2 years ago

            I was questioning your assertion that “It’s almost entirely an AI training card” so wondering whether it also make sense for people that just want the DP performance?
            I’m not up on Nvidia’s pricing for cards with decent DP compute performance.

            • MathMan
            • 2 years ago

            It is their best FP64 GPU since the Tesla K80.

            And it is the first Titan with decent FP64 since the Titan Z.

            • smilingcrow
            • 2 years ago

            So also a good value card for DP if you can live with the relatively low amount of memory.

        • Bensam123
        • 2 years ago

        It’s a Titan and not a Quadro or whatever Nvidia names their workstation class cards right now. I assume they’re remotely trying to target gaming enthusiasts… More so if this card turns out to be a turd because it’s just designed for compute workloads it’s going to cause quite a bit of problems as well (people buying it because they think it’s a halo product), as all past Titans have generally been the ‘best’ Nvidia GPU you could get for gaming.

          • Redocbew
          • 2 years ago

          Yes, I’m sure all 6 of them would be heartbroken.

            • DancinJack
            • 2 years ago

            You know, ironically, I think they may actually sell quite a few of these given its more compute focused capabilities relative to the Titans that have been on market recently. So like, AT LEAST 16 or 17 this go round.

            • Bensam123
            • 2 years ago

            You should visit Twitch more often. Even outside of it, they’re considered the ‘crown’ of gaming performance. Why do you think they have the Quadro name brand?

    • Wren
    • 2 years ago

    As far as I know, the full chip actually has 84 SMs for 5376 CUDA cores. Also the disabled L2 indicates that some ROPs are also disabled along with the 4th HBM stack and associated memory channels.

      • jihadjoe
      • 2 years ago

      Can’t be easy getting a full 800mm²+ chip. These GV100 chips that end up in Titan Vs are almost certainly those that got errors in critical portions that disqualify them from being sold as Teslas.

    • Ninjitsu
    • 2 years ago

    So is no one going to acknowledge that the name is a sorta pun on the Saturn V and Titan IV rockets?!

      • ArdWar
      • 2 years ago

      Because it would make a weak pun. What if it is a pun of Atlas V? or Minotaur V? or even Ares V? After all they all are of Greek mythology just like Titan, unlike Saturn which is Roman.

      The fact that last few member of Titan rocket family were one of the most overpriced launch system ever operational is interesting though….

        • Ninjitsu
        • 2 years ago

        But Titan is a moon of Saturn 😛
        And well, the Titan IV was an actual thing…

          • Mr Bill
          • 2 years ago

          I get it! Titan V is that part to fix the spaceship in ‘The Sirens of Titan’.

    • Mr Bill
    • 2 years ago

    Tensorlation benchmark anyone?

      • DancinJack
      • 2 years ago

      Nice.

    • Cuhulin
    • 2 years ago

    Can I use two of them with SLI?

      • tsk
      • 2 years ago

      No this card doesn’t have SLI connectors on it, it has NVlink connectors instead, those are disabled and you’d have to buy the quadro card for that.

    • Aquilino
    • 2 years ago

    So, next year still won’t get stable 16.7ms at 4K in Quantum Break.

      • DancinJack
      • 2 years ago

      For the 27 people that bought that game, it’s a real shame, I know.

        • Aquilino
        • 2 years ago

        Yes, please, look at the finger.

      • mcarson09
      • 2 years ago

      The steam version runs fine!! It’s the losers that paid for the windows store version that got burned. The truth is Microsoft doesn’t care!

        • Prestige Worldwide
        • 2 years ago

        I’m planning on picking it up once it hits $5

    • wingless
    • 2 years ago

    Looking at these comments, Nvidia’s marketing is working on the normies marvelously.

    I’m an Nvidia users but even I know that 110 ‘Tensor tflops’ DOES NOT EQUAL 110 gaming tflops. This GPU will be a little faster than a 1080 Ti in gaming workloads. It’s just an evolutionary performance jump when games are involved.

      • chuckula
      • 2 years ago

      [quote<]I'm an Nvidia users but even I know that 110 'Tensor tflops' DOES NOT EQUAL 110 gaming tflops. [/quote<] Show me one single comment anywhere here that even remotely implies that this card is some kind of gaming monster*. You won't find one because other than you nobody has said that, and in fact myself and others have literally said the exact opposite in multiple posts. * Well, it *is* a gaming monster due to its sheer size, but it's not worth it for $3000 compared to actual gaming cards since the features that make it a $3000 card have nothing to do with gaming.

      • DancinJack
      • 2 years ago

      To which comments are you referring? I don’t see any that would represent what you claim.

        • K-L-Waster
        • 2 years ago

        Oh, c’mon guys, let the AMD fan have his fun with his strawman.

          • Klimax
          • 2 years ago

          That can have nice alternate interpretation…

    • Mr Bill
    • 2 years ago

    Things are getting tensor between AMD and NVIDIA.
    This puts a whole new vector on the video card market.
    .
    .
    .
    .
    I’ll show myself out…

    • Chrispy_
    • 2 years ago

    Champagne.

    That’s gonna be awkward for the RGBLED crowd 😛

      • UberGerbil
      • 2 years ago

      Because… none of them are old enough to drink?

        • mcarson09
        • 2 years ago

        Comment of the week!

      • Anovoca
      • 2 years ago

      C9B6A5 – you are welcome kids

      Side note, this cooler will make the Noctua fan crowd very pleased.

    • drfish
    • 2 years ago

    I was beginning to this this wouldn’t happen.

    Now, the real question is, if they make a Volta Ti part, what will they call it? 1180 Ti or 2080 Ti? Taking all bets…

      • Bumper
      • 2 years ago

      Straight to 20. Nvda dont care. Leather jackets and big silicone.

      Yep.

    • ronch
    • 2 years ago

    Any hope for AMD to catch up?

      • DancinJack
      • 2 years ago

      At what?

        • ronch
        • 2 years ago

        Performance? Energy efficiency?

          • DancinJack
          • 2 years ago

          I mean, there is always hope. But it will take quite a bit of time and money for AMD to catch up on perf/watt. Also a better/newer process. I don’t think anyone is expecting Navi to be parity on perf/watt with Volta.

      • chuckula
      • 2 years ago

      Volta is just a shameless ripoff of innovations that AMD has already put in Navi, so clearly it’s Nvidia that has to catch up.

        • ronch
        • 2 years ago

        So is Navi out already?

        Edit – I think you’re being sarcastic.

          • K-L-Waster
          • 2 years ago

          [quote<]I think you're being sarcastic.[/quote<] What, Chuck, sarcastic? Nah, that's crazy talk....

            • ronch
            • 2 years ago

            Sorry, I’m new here…

        • freebird
        • 2 years ago

        And setting mining speed records for Ethereum… 77-80 MH being reported, but even with the elevated price of Ethereum, is will take awhile to payoff a Titan V, which can net about $180 a month at current pricing levels. Titan V only has 653 GB/s of memory bandwidth, so it will be interesting to see the mining speed of a card with more memory bandwidth.

      • Prestige Worldwide
      • 2 years ago

      It does not look promising, but we can always hope for them to pull another RV770.

    • the
    • 2 years ago

    This is stupid crazy powerful at a stupid crazy price… and I want [i<]two[/i<].

      • Jeff Kampman
      • 2 years ago

      Can I borrow one when you pull the trigger? 😉

        • DancinJack
        • 2 years ago

        I’d like the other. Thanks the (I think you’re in KS too so you can just drop it by my place kthx)

        • the
        • 2 years ago

        Well considering that this card [i<]doesn't[/i<] support nvLink or SLI, the second one would be a paper weight. Guess wait for the Quadro GV100 then?

          • ptsant
          • 2 years ago

          You can run calculation and deep learning on as many as you want in parallel. If you want to game, your money is probably better spent on the 1080Ti…

            • the
            • 2 years ago

            True but nvLink does permit a higher bandwidth and presumably lower latency coherency path between cards. Equally important is that memory from one card can be directly accessed by the other through nvLink without having to invoke the host system. This helps scaling up a system.

            Long term, I do think nVidia will simply go with [url=http://research.nvidia.com/publication/2017-06_MCM-GPU%3A-Multi-Chip-Module-GPUs<]multi-die GPU on an interposer[/url<] (or similar like EMIB) to combat yields and the slower rate new process nodes are arriving. Once that happens, cards like the Titan V will seem pedestrian in comparison.

        • tipoo
        • 2 years ago

        I suggest we launch plan get Jeff a Titan V to test. I…Uhh…Started it, so someone else say the next step.

    • techguy
    • 2 years ago

    Everyone thought $1200 was outrageous last year when Nvidia released the 2016 Titan X (Pascal). Welcome to 2017, and I would just like to congratulate AMD for failing so utterly to compete at the high end of the GPU market, thus causing this situation.

    We can only hope next generation Geforces don’t also see a price increase, but there’s no reason to think they won’t at this point. If AMD’s newest GPUs barely compete with Nvidia’s 18+ month old GP104, GV104 isn’t going to be pretty from a price perspective.

      • chuckula
      • 2 years ago

      Aside from fanboys who complain when Nvidia charges anything for any product, there’s nothing particularly “outrageous” about $3000 for a product like this given its intended niche.

      Adjusted for inflation it’s [b<]cheaper[/b<] at launch than the Titan-Z was: [url<]https://techreport.com/news/26520/nvidia-geforce-titan-z-is-selling-online-for-3-000[/url<]

      • MathMan
      • 2 years ago

      Nvidia never released a Titan version of P100.

      I prefer the option of choosing not to buy a Titan V100 over not having a choice at all.

      A lot of people will be very happy about this product.

      • Jeff Kampman
      • 2 years ago

      This is a more versatile and capable compute product than the Titan X Pascal, so the much higher price shouldn’t be that surprising. The vast and nebulous field of AI has a lot of dollars in it right now, and this card can do some AI-specific accelerations alongside every other form of GPU compute you might want.

        • techguy
        • 2 years ago

        My complaint stems from the fact that Nvidia insists on being ambiguous about Titan’s market position. Is it a high-end gaming card? Is it a professional card?

        With Pascal generation the argument that it is a professional card was a very hard sell. The first generation Titan X (Pascal) at least had more VRAM than any Geforce at the time, but by the time Titan Xp launched 1080 Ti had also released and with only 1 fewer GB of VRAM, while offering 95% of the performance at 58% of the price. Furthermore, Pascal generation Titan X cards offered *no* meaningful additional features or performance in professional workloads due to lack of compute differentiation compared to Geforces (i.e. FP64).

        I will admit that the Titan V is a better proposition as a professional card – without hesitation. Actually offering usable FP64 performance (and access to tensor cores) makes this very attractive to a variety of end users. Gamers though, aren’t among them.

        And therein lies the rub. Titan has always existed in this amorphous market that Nvidia likes to avoid defining so they can market it to anyone and everyone as “the best card you can buy” regardless of whether or not it actually is. Will Titan V be a faster gaming card than 1080 Ti or Titan Xp? Sure. Is it 3-5x faster in gaming workloads? Not a chance. 40% faster, tops. Will it be faster than “1180 Ti”? Not by much.

        My point then, can be summarized thusly:

        If you are a gamer, buying this card would be a mistake.
        Professionals with a use case for tensor cores and/or good FP64 performance *and* CUDA support: you may want to look at Titan V.

          • dragontamer5788
          • 2 years ago

          NVidia’s marketing seems pretty clear cut to me.

          * Server-rooms buy Tesla
          * Professional Workstations buy Quadro
          * Consumers buy GeForce, including Titan.

          The Titan line has always been for people who like shiny things and spending more money than the next guy. Its a solid market segment. Serious professionals need the better driver support (either from Quadro graphics for CAD work, or CUDA / Virtualization drivers associated with the Tesla cards)

          • DoomGuy64
          • 2 years ago

          It is a mistake to buy this as a gaming card for [i<]normal[/i<] people. The problem is that the PC gaming demographic is MASSIVELY tilted towards rich PCMR epeen fanboys, who WILL spend $3K on a graphics card, and $800 on a TN panel Gsync monitor, then tox up the forums like chuckula, saying how wonderful their shiny new Titan is. Nothing wrong with using it as a prosumer card that can also play games, but I wonder how many people will actually use it for that purpose. If not, PC gaming is dead. PCMR elitists killed it, and regular gamers will be relegated to consoles.

            • strangerguy
            • 2 years ago

            You are only offended because the rich and HPC guys are only buying NV and not from your favorite company. I’m sure AMD will never want to have a high-margin business like NV if they could because when they do they will clearly pick GPU social justice and their dreadful fanboys over profits, amirite?

            • DoomGuy64
            • 2 years ago

            No, I’m offended by toxic PCMR people in general, not who they’re buying from. The problem is that they act like there is nothing wrong with their behavior, toxing up everything, and they are also funding Nvidia’s predatory business practices. Nvidia is the perfect company for jerks, as they feed off one another.

            Nobody needs a $3000 Titan, or $800 TN Gsync monitor, but the Nvidiots sure think you do, as they preach their vendor lock in brainwashing propaganda everywhere. If anything kills PC gaming, it’s these morons. Consoles never were the problem, they were the solution to avoiding this toxic garbage.

            • DancinJack
            • 2 years ago

            lol

            • strangerguy
            • 2 years ago

            Yeah, it’s almost like NV put a gun on my head to buy their truly outrageously priced GTX 1070 and XB271HU.

            Have you AMD fanboys petitioned to your local congressmen yet about how there needs to be a right for cheap luxury goods because *~. ethics .~*?

            • DoomGuy64
            • 2 years ago

            Hey you know what? Enjoy your new Battlefront 2 p2w game over non internet net neutrality Comcast with your new Titan V and $800 TN panel Gsync monitor.

            The same people who complain about NN sponsor Nvidia. Go figure. Hypocrites.

            • K-L-Waster
            • 2 years ago

            At what point does this become an R&P thread?

            • Redocbew
            • 2 years ago

            I’m getting more of a “high school cheerleader” vibe where there’s random hysterics aimed at nothing in particular, but those hysterics are clearly very big and extremely important.

            I suppose one could turn into the other though.

            • Chrispy_
            • 2 years ago

            I love how you’re getting so riled up about this. The V100 isn’t even aimed at PCMR, Neither really is the Titan range as a whole.

            If you want to get all angry at PCMR, wait until the Geforce variant comes out and then you’re welcome to froth at the mouth all you want. I may even join you if I’m outraged enough by Nvidiots spouting brainwashed FUD all over the web, but that day is not today. Wait until Geforces launch.

            • Prestige Worldwide
            • 2 years ago

            Take an Ativan and a few deep breaths, my dude. Then keep calm and un-bunch thy panties.

            • K-L-Waster
            • 2 years ago

            If you think NVidia cards offer such bad value maybe you shouldn’t buy one.

            • Waco
            • 2 years ago

            Normal people can have nice things too, and these fit the bill for someone who wants the power of a Quadro without having to pay the price for one…oh, and it can game well. Refraining from insulting fellow members (especially unjustified insults) will probably help you keep that negative score a little closer to neutral. 🙂

          • psuedonymous
          • 2 years ago

          [quote<]My complaint stems from the fact that Nvidia insists on being ambiguous about Titan's market position. Is it a high-end gaming card? Is it a professional card?[/quote<]It's whatever the currently available chip does well. Titan through its history has been BOTH a 'better than the top gaming card but sucks for compute' card, AND a 'great for compute but you can get a better gaming card for less' card. Engage brain before purchase.

        • ptsant
        • 2 years ago

        Several people in our lab have bought the Titan X for AI. In fact, for research the Xp has been very popular. I expect the Titan Xv to do well in that setting.

      • K-L-Waster
      • 2 years ago

      The previous Titan was basically a halo-product gaming card. Think of it as a Corvette ZR-1.

      This is a professional card for compute workflows and not really intended for gamers. Think of it as a Corvette GTLM race car — if you’re not entering in the Daytona 24 Hours, you’re not in the market for it.

      (Edit: fixing dyslexia in my IMSA categories…)

      • Kougar
      • 2 years ago

      You seem to be ignoring that this amounts to a Tesla card at 1/3rd the cost. Instead of viewing it as an overpriced consumer card it should be viewed it as a deeply discounted Tesla card.

      • Meadows
      • 2 years ago

      This product is not relevant to the consumer market in any capacity.

    • Krogoth
    • 2 years ago

    How much can it fold?

    I wouldn’t be too surprised if it blows away AMD RTG GPUs at crypto-mining.

      • chuckula
      • 2 years ago

      [quote<]How much can it fold?[/quote<] Lemme check. Three towels and two dress shirts per second.

      • the
      • 2 years ago

      You could ask [url=https://www.youtube.com/watch?v=KuG_CeEZV6w<]these guys[/url<] and get an answer.

        • Mr Bill
        • 2 years ago

        I LOL’d

        • Mr Bill
        • 2 years ago

        Pretty sure this is a materials problem brought on by the fact that most paper while made of cellulose also has clay to improve the writing surface. One or the other reached a density difference that caused the mixture to separate and fragment.

      • dragontamer5788
      • 2 years ago

      Those “TensorCores” are the main feature, and I don’t think they will be able to do Folding@Home or Cryptomining.

      “TensorCores” are FP16 acceleration machines with a really weird architecture. They are specifically designed for matrix multiplications that are common in Convolutional Neural Networks, but I’m not sure if that specific matrix-multiply is found in many other problems. Unless I see some papers where other people find uses for the TensorCores, assume those things to be “TensorFlow” or “Caffe” specific. (They use half-floats, 16-bit floats, because speed is more important than accuracy in Convolutional Neural Networks). Its really hard for me to think of other tasks that would benefit from lots and lots of 16-bit FP matrix multiplications.

      With that said, I noticed that in [url=http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#maximize-instruction-throughput<]PTX 7.0[/url<] Integer Bit-shift is the same speed as 32-bit Floats now, so Volta should be faster in crypto-mining now. Volta == PTX 7.0. Consumer Pascal == PTX6.1, Server Pascal == PTX6.0. Notice bitshifts were 1/2 speed of 32-bit floating-point in 6.0 Pascal, but are now full speed in 7.0.

        • Crayon Shin Chan
        • 2 years ago

        For which mining algorithms does bitshifting matter?

          • dragontamer5788
          • 2 years ago

          I’m not an expert on mining. So someone else should answer this. Here’s my best guess.

          * Bitcoin is SHA256, it has a lot of bit-shifts but it doesn’t matter cause everyone seems to be ASICs now.

          * LTC is SCRYPT, which has a lot of bitshifts as seen in its [url=https://tools.ietf.org/html/rfc7914#page-4<]algorithm[/url<]. I don't know if LTC coins have gone ASIC yet. * Monero is a combination of hashes, but one of which is "BLAKE", which is [url=https://en.wikipedia.org/wiki/BLAKE_(hash_function)<]documented here[/url<]. * Etherium is Keccak-256 variant, incredibly similar to SHA3 but it has some differences. Keccak uses "Rotates", which are commonly implemented as bitshifts on GPUs. (CPUs have rotate instructions, GPUs... maybe? Depends?) Anyway, to "rotate" you bitshift left AND right and then "OR" the numbers together. And that's that. Bitcoin, Litecoin, Monero, and Etherium all use "bitshifts" significantly in the inner-most loop of their mining algorithms. I'd imagine that virtually every cryptocoin uses bitshifts a lot, but again, not an expert.

          • Klimax
          • 2 years ago

          Almost all of them. And not just bitshifting but rotating too. (Strangely, not even AVX2 includes vector rotating, only bit shift)

        • mudcore
        • 2 years ago

        Pretty curious to see how it fares for mining. Could take a bit to realize the full potential. It was a few months before the Vega GPUs were really tapped to their potential. This could mean an even bigger adjustment period. Though at $3000 a pop they really would have to be a whole new world of performance/efficiency.

          • K-L-Waster
          • 2 years ago

          You could buy 4 Vega64s for the price of this (if you can find them in stock of course) so I’d be surprised if many miners jump on it.

          Of course, miners have surprised me before, so who knows…?

            • Blytz
            • 2 years ago

            And then heat your home with them.

            • dragontamer5788
            • 2 years ago

            [quote<]You could buy 4 Vega64s for the price of this (if you can find them in stock of course) so I'd be surprised if many miners jump on it.[/quote<] Miners care about performance per watt. If Volta is more efficient than Vega, then the ever increasing electricity costs of mining would favor Volta. Of course, Vega may be more efficient than Volta in cryptomining. So it fully depends on the cost of electricity, as well as how well optimized the mining algorithms are for calculations per watt.

      • NovusBogus
      • 2 years ago

      It will be interesting to see the performance per dollar figures, since these things are used for the sorts of tasks where 3-4 GPUs aren’t really at a disadvantage compared to just one.

      • tipoo
      • 2 years ago

      The tensor performance is the main boost here, CUDA core performance is only up a bit in comparison. It would probably be a good mining card if there was no price to any cards since it’s a bit faster and the bandwidth is nice, but two 1080s would be much better anyways. So this won’t be a popular mining card, it’ll be a popular NN training card.

    • USAFTW
    • 2 years ago

    GOLD! (or beige?)
    I will have you… eventually.

    • kcarlile
    • 2 years ago

    Who knew that 4GB of HBM and a little bit more copper for a heatsink cost $7000…. Thanks, nvidia… (in re V100 vs. Titan V)

    • chuckula
    • 2 years ago

    At first I was going to laugh at yet another Pascal high-end rebrand, but $3000 for a full-bore Volta is actually pretty reasonable considering what Nvidia charges in this product range.

    I’d still never recommend it for actually doing graphics in a million years though.

    Oh and #PoorVolta was close but not quite right. It should have been more like: #VoltaWillMakeYouPoor.

      • K-L-Waster
      • 2 years ago

      Assuming that a GV104 version of this will have a similar MSRP to the 1080, you should be able to buy 4-5 GTX 1180s for the price of this (in 3-4 months…)

        • chuckula
        • 2 years ago

        As I said, I’d never recommend the Titan V for doing actual graphics. That’s what the purported GTX-1180 would be for.

        It probably won’t use HBM2 and there’s a good chance it will ditch FP64 performance for improved standard graphics performance too.

      • USAFTW
      • 2 years ago

      I agree, especially considering the significant upgrade in FP16, FP64 and tensor operations. Also, it is the largest processor ever created (815mm2, and on 12nm ff noless) with 4 (3 in this) stacks of HBM2.
      Considering the Titan Xp lacks many of the compute features in this, this is actually good value for prosumers.

        • smilingcrow
        • 2 years ago

        It’s much more of a card for Professionals than Prosumers and priced accordingly.

      • maroon1
      • 2 years ago

      Can you tell me which pascal card uses 12nm, and have 5120 cuda cores and has 640 tensor cores ?

      Rebrand is renaming same GPU to another name (like giving gtx 1080 ti a new name like gtx 2080 ti). I don’t see any of that with titan V

      Also, unlike Titan Xp. Titan V has full FP64 performance, support fp16, and has tensor cores. It is more expensive for this reason. If you do deep learning, then titan V is not expensive because it several times faster than any other GPU when comes to that

      Gaming performance probably won’t be much better than Titan Xp. Maybe 20% at most

      The gaming version of this card (maybe V102) will has much smaller die size, because they will cut down FP64, FP16 and also tesnor cores. It will also clock higher at same 250w TDP

        • NTMBK
        • 2 years ago

        This is a rename of the Tesla V100.

          • D@ Br@b($)!
          • 2 years ago

          Which is Volta architecture not Pascal

      • DoomGuy64
      • 2 years ago

      Lol, the original titan was nowhere near that. This is far beyond inflation. I think YOU should buy it though, all things considered.

        • NTMBK
        • 2 years ago

        Compare it to what you would pay for any other V100 or P100 GPU. This is the cheapest way to buy into V100 level deep learning performance.

        • chuckula
        • 2 years ago

        As I said earlier, the last compute-focused Titan, the Titan-Z, was also $3,000 in 2014.
        But expecting you to process easily verifiable factual information regarding Nvidia is clearly a lost cause.

      • D@ Br@b($)!
      • 2 years ago

      When used for the right workload/task:
      #VoltaWillMakeYouMillions

Pin It on Pinterest

Share This