Nvidia just posted this on Twitter: This is TITAN X

[Meladath]

If you want to practice doing high-end CUDA algorithm development, you'd much rather pay $1000 for the top Stream Processor count available than $3000 or $4000.

 

Or you know, you could just do OpenCL development, which is just as fast, and buy any AMD card instead, since they don't butcher the DP performance of their card anywhere near as much as Nvidia do to artificially reduce TDP numbers.

[patrickjp93]

The best part is people will actually buy this.

 

Its so hilariously pointless. It's just a 980Ti (or whatever they call it) with twice the VRAM and less butchered DP support. Except it has no ECC.... anyone working with DP will want ECC, and pretty much every application of DP will want people to use a quadro ANYWAY just for the professional drivers, which is what you pay for when you get a quadro.

 

The only people who buy titans are ignorant or just have too much money, there is always a better alternative.

 

It's even funnier too because nvidia's marketing team has done a GREAT job of fooling people into arguing over "its a gaming card, its not a gaming card" and somehow avoiding the fact that this is literally the most overpriced thing EVER. It is a gaming card, since it has no ECC and no professionally certified drivers, they even market it as a gaming card, but also market it as a "professional GPU" on the side in places that informed people look, in order to put two opinions across, so people argue over it! It's a genius work of social engineering thats for a fact.

 

AMD cards have less butchered DP support on pretty much all their cards, do that somehow mean that those cards are now "professional" GPU's (but not really, we market them as gaming, but not really gaming! ARGUE OVER IT BUT JUST GIVE ME MONEY)?

 

NO.

Maxwell is not native DP capable. anandtech and others have covered this before many times. You're stuck with 1/32 DP support no matter what card you get. Maxwell does not have 64-bit stream processors. It can only use mantissa extension. This is why kepler 200 chips are hitting the market on the quadro/Tesla side right now. There won't be a new workstation chip until Pascal.

 

Actually you pay for ECC support. The drivers are not very different from the Geforce ones if you actually bother to break into the code (disassembler).

 

It's a CUDA research card more than anything else, because games don't actually need that much RAM (see my previous comments because I'm not fleshing this out again), while scientific computing and rendering do, but if you're not a full-blown professional studio with tons of money, then a Titan worth $1000 is much more affordable than the equivalent Quadro/Tesla at 3x or 4x that price.

 

[TechnoBacon]

I may refer you to jayztwocents rage video

[patrickjp93]

Or you know, you could just do OpenCL development, which is just as fast, and buy any AMD card instead, since they don't butcher the DP performance of their card anywhere near as much as Nvidia do to artificially reduce TDP numbers.

OpenCL doesn't perform as well, though it is an open standard. Also, we're discussing the Titan line specifically which has full DP support, not 1/32 like Nvidia's other Geforce cards.

I'm evaluating purely on its merits and intended purpose. A lot of applications still use CUDA rather than OpenCL and croak when trying to use AMD GPUs. It's just life.

[Meladath]

Maxwell is not native DP capable. anandtech and others have covered this before many times. You're stuck with 1/32 DP support no matter what card you get. Maxwell does not have 64-bit stream processors. It can only use mantissa extension. This is why kepler 200 chips are hitting the market on the quadro/Tesla side right now. There won't be a new workstation chip until Pascal.

 

Mind telling me how the titan will magically have far better DP performance then if maxwell is not native DP capable and thus make it useful as you said?

 

I was more ranting for both the original titan and though less so than the titan X specifically, kepler clearly was capable of DP performance.

[patrickjp93]

Interesting , the 395x2 will probably cost less and features 8 thousand stream cores.

 

And then the dual-chip Maxwell Titan will follow with 6144 cores, and this time maybe under water as well, which will leave AMD far less performance breathing room.

[AxelRantila]

I may refer you to jayztwocents rage video

What did he say/do?

[Meladath]

OpenCL doesn't perform as well, though it is an open standard. Also, we're discussing the Titan line specifically which has full DP support, not 1/32 like Nvidia's other Geforce cards.

I'm evaluating purely on its merits and intended purpose. A lot of applications still use CUDA rather than OpenCL and croak when trying to use AMD GPUs. It's just life.

 

There have been quite a few studies showing OpenCL has very similar performance to CUDA, sadly the newest one appears to be quite a few years out of date.

 

I am evaluating it on the fact that its hilariously overpriced and justified in the dumbest possible ways.

[Guest]

Except it doesn't hold anything back. The 780 and titan were within a month of each other. Same with the TI and TB. The thing is the titan gets DP and extra memory, something that at the time no games needed (if you're using 4GB of RAM at 1440p or lower the developers are doing it wrong and you can prove this with memory use profiling. If the entire framebuffer is filled up but only 25% of it is being addressed in the span of a minute, then it's loaded down with junk it doesn't need to hold at that point in time). Scientific computing workloads are generally the only ones which use that much memory, and with good reason. A lot of optimization work goes into the algorithms in addition to having huge data sets to deal with. The same is true of rendering. Gaming very rarely is so high performance as to actually NEED more than 1.5-2GB in the frame buffer at a time even at 1440p +4x AA.

 

Just because the game uses it doesn't mean it should or even does so optimally.

 

Whether or not game developers are doing it wrong is not really relevant, some games simply "need" more than 3GB of VRAM to run properly. "Don't buy these games" is not really an option and it's not something that you want to hear when you just spent 500 dollars or more on a high-end GPU. Not to mention that SLI severely suffers here since you can use some crazy settings and forms of AA with the extra horsepower of one or more additional cards.

The Titan is not made for scientifc computation or even dp rendering, it has no ECC ram and you're stuck with the standard geforce drivers. It's not a proper workstation card, it lacks tons of features, software support and the proper drivers (+ ECC ram of course) to actually be used as a workstation card. Many AMD cards actually have fairly good double precision performance as well (because AMD doesn't aritficially gimp them on the consumer cards), I can't remember the exact numbers but they might even be comparable to the Titan, but do we call them workstation cards because of that? No, of course we don't. The whole argument is silly, the titan is and has always been a gaming card.

[digitalnav]

I'm going to benchmark 4 of these in SLI and wreck everyone's mind. How does that sound?

[patrickjp93]

Mind telling me how the titan will magically have far better DP performance then if maxwell is not native DP capable and thus make it useful as you said?

 

I was more ranting for both the original titan and though less so than the titan X specifically, kepler clearly was capable of DP performance.

It won't perform better. This is the 980TI/1080 flagship, not a separate thing this time around.

 

Only the original titans fit the bill of my description of the use and differentiation of the titan line. The Kepler Titans did nothing to hold back gaming. Any claims to the contrary are bald-faced irrational raging not grounded in understanding of reality.

[Zeruel]

I'm going to benchmark 4 of these in SLI and wreck everyone's mind. How does that sound?

 

 

Please be a boss and do this!

 

Not only for 4k but 4k surrounds and maybe the ROG Swift 144Hz.  

[TechnoBacon]

What did he say/do?

Watch it tiger

[zappian]

And then the dual-chip Maxwell Titan will follow with 6144 cores, and this time maybe under water as well, which will leave AMD far less performance breathing room.

 

Last time the dual chip battles happened the 295x2 won over the titan Z easily.

Considering a price/performance standpoint or just a performance standpoint in SOME situations.

And that is just a fact.

[Meladath]

It won't perform better. This is the 980TI/1080 flagship, not a separate thing this time around.

 

Only the original titans fit the bill of my description of the use and differentiation of the titan line. The Kepler Titans did nothing to hold back gaming. Any claims to the contrary are bald-faced irrational raging not grounded in understanding of reality.

 

Surely then, that means that this Titan X is even MORE overpriced than the original? Given that it will literally just be a 980Ti with double the VRAM? It has no ECC, no professional drivers, no real DP support?

 

EDIT:

 

https://www.youtube.com/watch?v=jJfICfX2-_0

 

This is the jayztwocents rage video previously mentioned

[-CLASSIFIED-]

Please Nvidia , make this affordable .... Because this will now be my next gpu lol

[s3ns3]

go fuck yo self nvidia with all your 1000$ titans.

 

it would be so stupid to release it before amd makes a move. couse then amd will again release a card thats better than this titan x at 1/3 of the price,ppl will bash nvidia for it,then nvidia will release the 980 ti and later titan XXX reloaded ,fully unlocked gm200 for just 1500$,because the 980 ti will perform better than the titan X

 

couse aparently the gtx 980 is way too cheap

[patrickjp93]

Whether or not game developers are doing it wrong is not really relevant, some games simply "need" more than 3GB of VRAM to run properly. "Don't buy these games" is not really an option and it's not something that you want to hear when you just spent 500 dollars or more on a high-end GPU. Not to mention that SLI severely suffers here since you can use some crazy settings and forms of AA with the extra horsepower of one or more additional cards.

The Titan is not made for scientifc computation or even dp rendering, it has no ECC ram and you're stuck with the standard geforce drivers. It's not a proper workstation card, it lacks tons of features, software support and the proper drivers (+ ECC ram of course) to actually be used as a workstation card. Many AMD cards actually have fairly good double precision performance as well (because AMD doesn't aritficially gimp them on the consumer cards), I can't remember the exact numbers but they might even be comparable to the Titan, but do we call them workstation cards because of that? No, of course we don't. The whole argument is silly, the titan is and has always been a gaming card.

 

Not buying those games is always an option, or are you willing to give up your power as a consumer and pay more than you should have to for a given experience?

 

Not wanting to face reality is just the approach of the mentally immature.

 

The whole point is to let a budding programmer or startup studio do it cheaply (meaning you go without the insurance of ECC). It serves a niche but distinct and important purpose: lowering startup costs.

 

Instead of Teslas my school just tossed a pair of titan black cards into one of our server blades for the HPC and Heterogeneous Computing classes to use. That way we could experiment on a variety of program scales but not have to pay 3-4x more for the equivalent Quadro or Tesla cards.

 

Believe it or not if you dig around in the driver code the professional and geforce drivers are about 95% similar. Software does not rely on professional drivers. It only assumes they are there and are tuned to perform based on the existence of those drivers. What actually happens on the hardware is transparent to the program as long as outputs are correct.

 

Nvidia GIMPs to be able to run cooler, hence its popularity with ITX system builders.

 

the argument is not silly at all. Only your counterargument is.

[Rodneyg87]

This is completely incorrect for a number of reasons. The original titan and 780 released within a month or two of each other. The Titan is equipped with full DP support. The 780 has 1/32 performance support only. The Titan is meant to be a poor man's quadro capable of both scientific computing/rendering and gaming, but without the costly insurance of ECC memory. they're pretty much HPC research tools at a discount over the Tesla/Quadro lines. If you want to practice doing high-end CUDA algorithm development, you'd much rather pay $1000 for the top Stream Processor count available than $3000 or $4000.

 

Titans do not hold back gaming at all. In fact the very idea we need 4GB or 6+ for gaming is absolutely ludicrous. Run a memory utilization profiler on your GPU and watch how much of those 4GB are actually being accessed per minute despite being filled up. It's around 25-35% per minute for modern AAA games. between 1 and 2GB even in 1440p with 4x AA. That should tell you something about bad design and memory management. Scientific computing and rendering actually use all of that memory clock by clock doing huge parallel transforms of the whole data set. Games right now leave a ton of textures and unnecessary data sitting around despite being nowhere near a part of the game which could use them.

 

Just as with people's complaints about Intel's "small" performance gains, quit your bellyaching. It's the fault of software makers for sucking (focusing on legacy support in the case of CPU programs and just bad overall design in GPU programs).

Your speaking in terms of something other than "gaming". My post was focused on gaming performance not scientific computing. And before you say that the Titan is not for "gamers" read the side of the box it specifically says "Inspired by Gamers" on the side which means the card is marketed for the gaming community as well, and if that is the case then my statement is true for the Gaming community and advancements in gaming technologies. I do however agree the card is as you said a "poor mans Quadro or Tesla" but they need to market it as that not as a gaming card then if it is its true intended purpose.

 

But by them marketing it as a gaming card it means that for the Maxwell generation the Titan will be seen as the pinnacle of performance and you will not see any card in the Maxwell lineup come close to its performance.

 

And as for the need for 2 gigs or 4 gigs of VRAM I could honestly care less because until I build my computer I am gaming on a Macbook pro with a GTX 750M with 2 gigs of VRAM and I can run heros of the storm, League of Legends, and hearthstone which are the only games I play currently until I finish my pc and I can play at 60+ frames per second at 2560x1800 resolution on high/ultra.

[IMPERIUS]

Sexy. Probably overpriced like the original titan was. I guess linus will be happy to get 10 of them sponsored from nvidia tho

[patrickjp93]

Last time the dual chip battles happened the 295x2 won over the titan Z easily.

Considering a price/performance standpoint or just a performance standpoint in SOME situations.

And that is just a fact.

 

It only wins because AMD put it under water. If you put the titan Z under an EK waterblock, guess which one actually wins in performance (and still in perf/watt)...

 

I stand by what I said AMD produced a failure in engineering and had Asetek bail them out. All things being equal the Titan Z can win, though its launch price was obnoxious and I'm not going to deny that.

[patrickjp93]

There have been quite a few studies showing OpenCL has very similar performance to CUDA, sadly the newest one appears to be quite a few years out of date.

 

I am evaluating it on the fact that its hilariously overpriced and justified in the dumbest possible ways.

 

If it was overpriced no one would buy it. Seriously people would it kill you to take a microeconomics class? Nvidia aims for equilibrium price like every other competitive company. That said, it targets a range of different demands and hits the equilibrium prices of the people in those various brackets. Just because it isn't right for you doesn't mean it's overpriced.

[zappian]

It only wins because AMD put it under water. If you put the titan Z under an EK waterblock, guess which one actually wins in performance (and still in perf/watt)...

 

I stand by what I said AMD produced a failure in engineering and had Asetek bail them out. All things being equal the Titan Z can win, though its launch price was obnoxious and I'm not going to deny that.

 

Oh yeah lets spend 3k on a titan Z at launch and then more money on a EK waterblock and a loop also.

I would get two titan blacks and water cool both and still save money compared to a titan Z.

Why the 1000 dollar premium?

[Meladath]

If it was overpriced no one would buy it.

 

Monster cables (or any other incredibly expensive digital cables), Beats by Dre, most audiophile equipment, apple accessories, diamonds and COUNTLESS other things that sell well would like to have a word with you.

[Opcode]

@patrickjp93 I had a feeling the TITAN X would make its way to the public before Bermuda/Fiji.