Slashdot Mirror
NVIDIA's GeForce GTX TITAN X Becomes First 12GB Consumer Graphics Card
Deathspawner writes: When NVIDIA announced its GeForce GTX TITAN X at GTC, no one was surprised that it'd be faster than the company's previous top-end card, the GTX 980. But what did impress many is that the company said the card would sport a staggering 12GB of VRAM. As Techgage found, pushing that 12GB is an exercise in patience — you really have to go out of your way to come even close.
Additional reviews available at PC Perspective and AnandTech. The latter notes, "...from a technical perspective, the GTX Titan X and GM200 GPU represent an interesting shift in high-end GPU design goals for NVIDIA, one whose ramifications I’m not sure we fully understand yet. By building what’s essentially a bigger version of GM204, heavy on graphics and light on FP64 compute, NVIDIA has been able to drive up performance without a GM204-like increase in die size. At 601mm2 GM200 is still NVIDIA’s largest GPU to date, but by producing their purist graphics GPU in quite some time, it has allowed NVIDIA to pack more graphics horsepower than ever before into a 28nm GPU. What remains to be seen then is whether this graphics/FP32-centric design is a one-off occurrence for 28nm, or if this is the start of a permanent shift in NVIDIA GPU design."
You sure it's 12Gb and not 11.5? ;-)
It's certainly a legitimate question.
On the other hand, as I remarked to a friend when discussing graphics cards the other day, it seems to me that none of the numbers on the cards really matter - RAM, Clock Speed, etc. There are so many variables that you might as well say that it has 12 GigaDrawing Cores and 256 SuperShaders with a 1.5 TeraTransit Hop Interval. What really matters is the final performance, using the universal standard measure of FoC(3), aka "Frames (per second) of Crysis3".
Honestly, when I look at graphics cards (and I'm currently doing so for my next system), that's what matters to me most in comparing them - looking at benchmark results among all the major games. Certainly Nvidia should be telling us up front how the architecture works, and that you have 3.5GB of main/fast memory or whatever, but benchmark statistics from 3rd parties seem to me to not only be more important, but not exactly (or at least not easily) fakeable.
That said, it's interesting to me that the benchmarks I've seen for this card mark it as inferior to the 295X2, which is not only almost a year older, but significantly cheaper. I'm personally hesitant to go with AMD, as my current experience with AMD drivers on my desktop has been downright horrible, compared to a relatively painless experience with my current laptop running an Nvidia card, but I have to wonder what's going on when Nvidia comes out with something like this, a year later, and at seemingly worse performance.
Deep Neural Nets will use it, at least the very big, wide, deep ones will.
@de_machina
In case people are confused it is important to point out that the Titan cards aren't aimed at gamers. They are partly a PR stunt for Nvidia (look, we make the biggest, baddest GPU out there), and partly of interest to developers working in graphics research (either developing tech for next gen games, GPGPU research, fluid simulations, and other projects). When you are raycasting massive voxel scenes for example, the 12GB can look rather attractive.
At the end of the day it is very much a niche product, and calling it a "consumer card" is perhaps a bit of a misnomer. If all you are looking to do is to consume content (i.e. play games) this isn't the card for you, just SLI as many GTX980's as you can afford together and be done with it.
I know some people who are drooling all over this card for various real-time simulations. The Unified Physics Solver and 12GiB RAM on the card will allow them to push a combination of good graphics and good enough physics, at decent frame rates in real-time, all on a single card. It also makes it easier to develop sim solutions.
The Titan cards ARE aimed at gamers.
Have you seen the reviews? Did you see the presentation today? Did you see the double precision compute performance? This is not a compute card.
Quadro = Workstation
GTX = Gamer
Titan = Gamer with more money than sense
In case people are confused it is important to point out that the Titan cards aren't aimed at gamers.
Yes, the Titan series is an odd bridge between consumer price/ performance, and professional reliability (ECC RAM) and unhampered double precision performance at painfully professional level prices of the CUDA / GPGPU oriented Quadro and Telsa cards. (Telsa K20 to K80 cost $3500 - $5000 USD approx AFAIK)
So they are great alternatives for CUDA aware 3D graphics application users who traditionally can't afford a Quadro or Telsa card (are not professional movie / video-game studio artists or CAD designers), and students and researchers looking for low(er)-cost prototypes for developing CUDA / GPGPU software / experience. I assume any graduate CS/EE student studying parallelism / many-core hardware or software wants a Titan as an affordable alternative to the tuition-like pricing of the Quadro and Telsa cards.
I had considering putting the Titan X on my own personal wish list, until I read that the FP64 isn't in line with performance improvements over previous generation Titans.
The reason Bitcoin was done on AMD GPU's was because AMD had a bitshift instruction, and that due to how simple Bitcoin is, AMD's rather retarded memory handling didn't cause any choking. Double precision had absolutely zero to do with Bitcoin running well on AMD.
AMD's memory handling means that the actual Double Precision performance you get out of a given card is extremely different from the theoretical maximum. In fact, for many real-world tasks, Nvidia+CUDA delivers better DP performance than the AMD paper tigers, especially per watt, and so the HPC world heavily leans towards Nvidia, even with DP heavy tasks. For tasks that doesn't need SP, like many forms of signal analysis etc? Forget AMD, not worth the bother of shitty drivers, piss-poor memory handling, worse performance per watt.
As for RAM, there are things it's useful for, especially since the Titan also has its Universal Physics Solver, meaning that together with the RAM amount, you can put together very detailed sims at decent resolutions and framerates, all on a single card.
Just to correct a misunderstanding many people have:
Not everyone who does compute needs double precision. Many tasks within signal processing only need single precision, for example.
Hell you can't even use Crysis 3 by itself as its been shown to run better than it should on some cards, worse on others. That is what happens when you run a poorly optimized game.
Frankly I wish more would do like Tek Syndicate (of course they aren't taking money in the form of ad contracts while many of them are getting huge checks from at least one) where they choose a half a dozen popular games that require some muscle and then see what you get, as IMHO this is a hell of a lot closer to what one would get IRL than "run one benchmark, run BF4" like the others tend to do.
But is there even a point in having 12GB of VRAM at this time? Most games aren't even hitting 4GB and by the time you have games hitting 12GB the GPU is gonna be so out of date it might as well be an FX5900, since its lousy FP means the CUDA guys aren't gonna want this,. So other than "look at my massive ePeen" is there a real use for 12GB now?
ACs don't waste your time replying, your posts are never seen by me.
I do wonder if ridiculous amounts of VRAM may end up being useful in game engines that are currently only on the horizon - for example, Outerra's 1:1 scale planet engine renders mostly in GPU. One wonders too about RAM demand of something like Euclideon's "unlimited detail" engine (assuming it isn't vapourware). If we're moving to games that do more in the GPU, then maybe stupid amounts of VRAM might actually get used? Then again, I expect if that were to happen (using complex GPU-based world generators), the compute performance is going to matter too...
I can still remember when I couldn't understand how you'd ever fill a 2GB hard drive... (this was before video on computers was a serious thing, obviously). 12GB VRAM is just bonkers.
Yeah I noticed that nVidia is playing games with the FP64 with the Titan X. :-(
Conflating high end gamers using Titan X with the scientific performance of the previous Titan Black and Titan is rather dodgy.
RAM is very important if you use high resolutions. If you game in 1080p, then yeah this won't tell you much. If you have a 4k monitor though, 3GB isn't enough so you can at least look at the RAM to narrow your selection, then look at benchmarks.
Religion is the best example of mass psychosis
I was at the keynote at GTC this morning and it really depends on what you are doing. If you want to do numerical simulation, it is not very useful because double precision performance is terrible. But if you do data mining, you mostly care about bandwidth and single precision performance. And then 12GB isn't too much. Actually I find it still a bit on the low side. Intel Xeon Phi are featuring 16GB this days. And in the realm of data analysis fitting the data on the accelerator is what make the difference ebetween the accelerator is great and the accelerator is useless.
I completely disagree. The majority of the HPC realm still uses Nvidia only because they know CUDA and not because of any technological advantage. AMD has held the line at not allowing sloppy programming methods into their OpenCL compiler and that has held back a lot of HPC users from jumping ship. You can even see this in many complaints from open source projects, like Blender, where they refuse to produce proper multi-threaded code and rely heavily on the CUDA compiler to do the work for them.
The rest of your complaints, "shitty drivers", "piss-poor memory handling" and "worse performance per watt" are also bogus. I own or manage machines using a large number of Nvidia and AMD video cards, and have seen as many driver issues between the two that neither has come out worse. This is a typical fanboy stereotype that keeps being repeated with no real fact behind it.
Your second complaint is seen a lot in programming forums, but I have never seen anyone do a proper write up of any memory issues with any of AMD's generations and most of the conversations lead me to believe it was an issue of the programmer's personal preference not wanting to learn a second platform with less market share than an actual technical issue. Most of these issues would be alleviate if the programmer would just use a common optimized library and stop trying to redo the work themselves.
Lastly, AMD's offerings have historically produce more performance per watt and their latest offerings continue that trend. This, besides the bit shift ability you mention, is also one reason why AMD was used for Bitcoin mining and supercomputers.
http://www.tomshardware.com/re...
http://www.green500.org/news/g...
Now, my latest personal computer has an Nvidia GeForce GTX 980 inside because I more often need to fix CUDA code and noticed some of the games I wanted to play ran better on it (again, from the game designer's preference and not a technical merit). I personally own eight other video cards across AMD, Nvidia and Matrox (who use AMD GPUs these days) and three generations for testing.
And I am only sticking up for AMD because I admire their push to get people to code for multi-core better. Nvidia has been too conciliatory in the last six years in that respect, which is fine for their revenue stream and market share but not a good thing in the big picture for the broader computer industry. Since Moore's law has begun to slow, we are going to need a massive shift to multi-core optimized applications and we need programmers ready for that day.
AMD seems to be ready with the tough love to get everyone there while Nvidia keeps enabling bad behaviors.