Slashdot Mirror
NVIDIA GeForce GTX TITAN Uses 7.1 Billion Transistor GK110 GPU
Vigile writes "NVIDIA's new GeForce GTX TITAN graphics card is being announced today and is utilizing the GK110 GPU first announced in May of 2012 for HPC and supercomputing markets. The GPU touts computing horsepower at 4.5 TFLOPS provided by the 2,688 single precision cores, 896 double precision cores, a 384-bit memory bus and 6GB of on-board memory doubling the included frame buffer that AMD's Radeon HD 7970 uses. With a make up of 7.1 billion transistors and a 551 mm^2 die size, GK110 is very close to the reticle limit for current lithography technology! The GTX TITAN introduces a new GPU Boost revision based on real-time temperature monitoring and support for monitor refresh rate overclocking that will entice gamers and with a $999 price tag, the card could be one of the best GPGPU options on the market." HotHardware says the card "will easily be the most powerful single-GPU powered graphics card available when it ships, with relatively quiet operation and lower power consumption than the previous generation GeForce GTX 690 dual-GPU card."
All games that have the budget for graphics these days are targeted at console limitations. I can't really see any reason to spend that much on a graphics card, except if you're a game developer yourself.
Hmm. $999 for 4.5 TF/s vs. $399 for 4.3 TF/s from AMD Radeon 7970. Hard to choose.
And here I was, thinking that TI-83 has pretty cool graphics.
Can I throw a bunch of these in my Bitcoin rig and pay back my intial fiat outlay with purestrain mined bitcoin? I don't see why that wouldn't work and could level the playing field against the ASICs that are coming online now and pushing older GPUs into the background.
Wow. 3x as many transistors as a Core i7 3960X? I guess the days are finally here when you buy your graphics card and then figure out what kind of system to add on to it, rather than the other way around.
"will easily be the most powerful single-GPU powered graphics card available when it ships"
Yep, for the first week or two. I'll stick with my 670 that runs BF3 at max settings with 50+ FPS. Graphics card like the Titan is as useless as Anne Frank's drumset for the typical gamer.
Software (other than games) that can actually benefit from this type of hardware is scarce and expensive. This $1000 card will probably be in the $5 bargain box at the local computer recycle shop before there is any significant software in widespread use that could put it to good use.
...so it includes support for modelines?
I have a 3d 120hz monitor, it would be nice to be actually able to play new release games at 120fps minimum with all the eye candy on, in at least 1920x1080.
I thought that most HPC users needed double-precision maths.
Why, then, would a card aimed at the HPC market have so many single-precision cores alongside the double-precision cores?
And, I'd say it's way overpowered. Right now, I can play BF3 and Eve simultaneously w/ no problems. I got it for future proofing my gaming needs. Hardware has to be ahead, though. If it wasn't gamers would be in a constant cycle of upgrading hardware. By getting the latest/greatest, I've seen that I can go about 5 years before needing an upgrade to stay current.
I swear to God...I swear to God! That is NOT how you treat your human!
My graphics card has a 16 bit pipeline!
and this card will run your current games at 200+ fps at 1080p while newer games will run at 40fps which means you will then have to buy a newer card. What is the wattage usage? 1000 watt psu minimum requirement.
What we need is gameplay which is missing in today's pc games this is why I always find myself going back to emulators. These day's pc's are consuming watts like air conditioners.
The software isn't coming because until recently it's been hard to program to take advantage of the hardware. When I can use Python to interact with this hardware then the software will come from people like me.
add more cores
How can we still claim to be building custom PCs when half the processing power is confined to a little sealed appliance? The way things are going PC gamers will just be putting custom faceplates on their AMD/NVidia consoles.
7.1B transistors in 551mm^2? That's atrociously low transistor density.
Most of us probably use things that are probably 1/8th the size with 16B+ transistors on it. You probably know them as little 32GB+ memory cards.
The thing is - memory devices (all memory - flash (NOR/NAND), RAM (SRAM/DRAM) etc) are the most transistor-dense things around - their sheer density makes it so that they're limited by how much silicon area they can use - if you double the silicon area, you double the storage. Moore's law helps here because you can stick twice as many transistors on for twice the storage, but the same silicon area.
Even silicon area isn't that impressive - a good dSLR will have a camera sensor with a large silicon area. Hell, there are FPGAs with just as big silicon dies as well.
In fact, that chip, the vast majority of those 7.1B transistors will probably occuply less than 25% of the entire area - being used for onchip caches and temporary buffers and memory, which are extremely dense structures. The rest of the area is taken up by very few transistors. Instead, what takes up the silicon area is... wiring. Even with 10 layer metal, there's miles of wiring inside the cihp - typically around the logic parts. The reason for this is logic is often called "random" because there isn't generally a regular rhythm to the blocks (sure each processor is regular, but within each one it isn't).
Finally - you can fab a chip up to the size of the water you use - it'll be uneconomic because every wafer is flawed, so the yield you get goes down exponentially as the size ramps up. Those FPGAs I mentioned? Easily $30K a pop. Yes, $30,000 for ONE CHIP. Because yields are horrendously low due to how big they are (if you're lucky, you'll get one good one per wafer. Which has to pay for all the bad ones).
Yield is a huge issue - the bigger the chip, the greater the chance of it encountering the imperfect part of the wafer, which cuts yield down. In addition, a larger die means you can fit less on a wafer, so you have less chips to pay for the fixed costs of a wafer and processing.
551mm^2 is probably the limit for economic production of this chip, but it's hardly stressing the technology.
Probably would not be able to run my $100,000 Triple display 85" 4k resolution computer screens so that I can play angry birds at 60fps.
Does it mention you can SLI these cards? Seriously I doubt one would be enough for my gaming needs.
Console rez means 1280x720, perhaps less. I know that in theory the PS3 can render at 1080, but in reality basically nothing does. All the games you see out these days are 1280x720, or sometimes even less. The consoles allow for internal resolutions of arbitrary amounts less and then upsample them, and a number of games do that.
Frame rate is also an issue. Most console games are 30fps titles, meaning that's all they target (and sometimes they slow down below that). On a PC, of course, you can aim for 60fps (or more, if you like).
When you combine those, you can want a lot of power. I just moved up to a 2560x1600 monitor, and my GTX 680 is now inadequate. Well ok, maybe that's not the right term, but it isn't overpowered anymore. For some games, like Rift and Skyrim, I can't crank everything up and still maintain a high framerate. I have to choose choppy display, less detail, or a lower rez. If I had the option, I'd rather not.
But what about for mining bitcoins?
Yes, I know that it's a dubious but it might make mining bit coins temporarily profitable again.
And this isn't meant as a troll comment, I'm just looking for opinion from the bitcoin community.
Finally, I can have my dream of more transistors on my graphics card than there are people on this planet. And since we used the advancing graphics card vs losing people way, no one had to die!
I believe there are two modern lithography lens manufacturers, one at 32mm x 25mm and the other at 31mm x 26mm, although I'm having trouble seeing publicly available information to confirm that. Either way, 800mm2 is the approximate upper bound of a die size, minus a bit for kerf, which can be very small. Power7 was a bit bigger. Tukwila was nearly 700mm2. Usually chips come in way under this limit and get tiled across the biggest reticle they can. A 6mm x 10mm chip might get tiled 3 across and 4 up, for example.
655,360 transistors should be enough for anyone.
Knowledge is how to play a game, intelligence is how to win, wisdom is knowing what game to play.
Hopefully their CUDA performance isnt crippled like it was on the 6xx series. The 580 has significantly better CUDA performance then the 680
UDL
I bet I will see these in our diagnostics workstations before long. The software we use for real time rendering/imaging is limited by the video card. If it will improve rendering/diagnostic times, we will end up buying them.
"...with a die size of 551 mm^2, the GTX TITAN is a big step forward (both in performance and physical size)."
Comes with it's own sub-chasis, I guess. .551 m ~= 1.8077 ft per side]
[551 mm =
Can someone clarify if the 4.5 TF is single or double? The hardware leads me to believe it is single.
4.5 TF is a huge number for SP, but unless I'm doing the math wrong, their DP flops are really low, less than 2 TLFOP. Single precision is great for games, but dual-precision is generally the supercomputer benchmark (LINPACK).
Thanks in advance.
https://www.accountkiller.com/removal-requested
I remember when the top-of-the-line cards were about $500 (and I never bought one of those either). Why do so when you can get 85% of the performance for half the price? Speaking of which, what's the current bang-per-buck darling in the nvidia lineup (around $250)? I currently have a GTX 560 ti, and I'm wondering if I could get a decent performance boost by upgrading.
Seriously man, this isn't a console-fan argument nor is that one you want to have in relation to PC hardware because you'll lose. The point is, most games these days are targeted at 1280x720, or lower, at 30fps. The problem is to target anything higher you trade something off. Want 60fps? Ok, less detail. Want 1080? Ok, less detail. There is just only so many pixels the hardware can push. Crank up the rez, you have to sacrifice things.
Computers can do more than that, but need more hardware to do it. The target on my system is 2560x1600 @ 60fps, with no detail loss. My 680 can't handle that all the time, that's the point.
Let's have a look at some recent non-FPS games:
Darksiders II: 1152x640
Dishonoured : 1280x720
Mass Effect 3: 1280x720
Need For Speed: Most Wanted: 1280x704
Soul Calibur V: 1280x720
Sleeping Dogs: 1152x640
X-Com Enemy Unknown: 1280x720
That's just a selection of games released last year, that aren't FPS's that use 1280x720, or lower, on the PS3.
Most PS3 games don't do 1920x1080. It doesn't have the fillrate, or the VRAM, to deal with that without some serious quality sacrifices so most developers choose less rez for more eye candy.
Remember that the resolution it is outputting at isn't the same as rendering. You can upsample any output you like, hence how a DVD player can output a 1080p signal though the DVD is 720x480 anamorphic.
They are aiming the GTX Titan at the GPU scientific computing market- by fitting up to three of the things in a smaller form factor and running on a (relatively) small power footprint of 250W, these can be useful for CFD (Computational Fluid Dynamics) calculations, quantum chemistry calculations, protein folding, bitcoin mining (as mentioned by msauve) the list goes on.
That's why it has the configuration it does (if some of the amounts it has look strange to you it is because there are units disabled) and why it costs so damn much. Combine a low number per wafer, with probable high failures (since TSMC's 28nm process has quite a few issues) and you get a high cost per working part, and thus a high cost to consumers.
will it play Crysis?
It is far and away the most popular game engine. However it is fully capable of 1080 rendering, or indeed far beyond that if the hardware can handle it. All those games go beyond 720 on the PC.
It is a tradeoff on the PS3. You can go to 1080 if you like, but that requires 2.25x the fillrate of 720. That corresponds to less complex shaders, less post processing, smaller textures (also because of less VRAM) and so on. Given that the Xbox 360 won't render at 1080, and that 720 is fairly "standard" for this gen of games, not to mention that it all looks better than the SD of last gen, most companies opt for a prettier picture with less pixels.
Same reason some games render even lower. Sleeping Dogs wanted more detail than they had fillrate for so it renders at a lower rez.
Sometimes I despair at how little some people understand of what they read before they post, or failing that if their attention span is limited to six words.
http://www.luxrender.net/en_GB/gpu_support