desperate for the fabs to get on the ball and beyond 28nm. And while 28nm is likely, its a good best guess, not yet confirmed, although Titan X debuting on something other than 28nm would likely be bigger news than the chip itself.
It has been talked about since last year. TSMC doesn't seem to be able to get higher powered processors on its 20nm process. It isn't clear if this was a decision on their part or that they found it just didn't work.
This has been known for about half a year now. GM200 is being fabbed on 28nm. What hasn't been known is the exact transistor count and die size. Rumors suggest it'll be the largest GPU Nvidia's produced, ~600mm².
My guestimation put it at 594 mm^2, pretty close. Remember that not all functional units are being increased (video decoder/encoder etc.).
Still not a record holder for chip size. Intel's Tukwila Itanium 2 was 699 mm^2. At the time, the 65 nm limit was estimated around 750 mm^2 due to the optics involved.
We need to ask if the VRAM is partitioned to two 6 GB sets or three 4 GB sets or one 12 GB. Fair question. TDP and price are the only items from the table above that I'm curious about. The each Titan chip has spent too little time as the top dog for the money they're asking. It maybe makes sense as a CUDA card, but it has none of the virtualization features.
It is total crap that 8 generations of GPUs need to be 28 nm. I totally understand it may not be AMD or Nvidia's fault. Too many processes are built for low power chips.
To be mentioned - the AMD R295 is now about $600 on Newegg. Don't forget it will cost $30 a month in electrical bills though.
If the number of ROPs is evenly divisible with the amount of memory, there doesn't need to be any partitioning. Its only a tool used when the numbers diverge, like 56 ROPs and 4GB.
However 96 ROPs on a 384-bit bus match perfectly with 6 or 12 GB of memory. Extrapolating from the 980, it will have 12 ports with a 32-bit MC each and 8 ROPs, each hosting 128MB of memory.
Okay so GM204 (980) has 64 rops. This has 96 (with a question mark next to it ) so thats 150% Adding a extra 50% to the other numbers from GM204 gives us: 3077 Cuda Cores 192 Texture Units 384 bit bus is 150% of 256% of the 980 This has 3 times the RAM, but it is a titan, so maybe 980ti will have 6gb (ie 150% of 980 again) Memory clock and core clock I would expect to be the usual. 7ghz and 1ghz-ish respectively. Oh and the 5.2 Billion transistors of the 980 * 150% would give 7.8, or close to 8 billion. So this times by 150% thing seem to fit pretty well.
(This is all from just noticing the simple 50% increase in ROPs and doing the numbers incase someone is interested, none of this is solid information).
I completely agree. These are the specs the leaks have pointed to as well.
I'll also note this: this will be a 24-SMM design, with all SMMs on the die enabled, as 7.8B is around the upper limit of what TSMC can put on a single die right now at 28nm.
This means there won't be a new version of this chip like there was with the original Titan and Titan Black.
I also expect there to be a consumer version of this with 22 SMM's enabled, with the requisite ROPs and L2 cache disabled like they were on the 970 It'll likely also have 6GB of VRAM (For yes, 5GB of fast VRAM or however the math works out for their bus partitioning). This Titan Lite or 980 Ti or 985 or 990 or whatever they're gonna call it will be the real card to get for 4K from NVIDIA until they go through a process shrink. Two of those will be quite potent.
The PCB will have 6GB, but it could be partitioned similarly to the 970's, which has a 3.5GB high-performance partition and a slow 512MB partition. I haven't worked out exactly what the equivalent would be based on SMM configuration, but it wouldn't surprise me to see 6GB on the PCB itself with something like a 768MB slow memory partition and 5.25GB fast memory partition.
Or they'll do something totally new. No idea really. It's hard to even speculate what they might do. For the 22 SMM, it's conceivable they'll partition the memory like on the 970, but for the 20 SMM version, they could just disable one of the L2/ROP/Memory Controller units and go with a 320-bit bus for that part with 5GB of VRAM. That would actually be a really good part to have between the 980 and the other GM200-based cards.
This is a "prosumer" card which is sold with the idea that it could be used for things other than games, hence the FP64 at 1/3 FP32 for the other two Titan cards. It's very unlikely they would do such memory partitioning since, although 12GB is not going to be used for gaming, it might be used in other situations.
The strange memory speeds in the 970 are due to the ROP set up now the SMMs, which can be completely independent. So you can have 22 SMMs and full speed memory. Overall the whole memorygate or whatever it is called was a bit of a mess, but as a technology the extra granularity is good for Nvidia (for yields) and us (they can be less conservative,ie 4gb of 3.5fast and .5 slo instead of just plain old 3gb). But I agree that while they are chopping off 2 SMMs for yields, they might mess around with the memory system again.
I've been hoping for a 8GB 980TI and this leaves them room for that. They're going to have to do that if AMD brings out a 8GB 390x, which they probably will.
The 20nm process isn't good enough for high performance GPUs, so the consensus is that there won't be 20nm dGPUs, but they'll stick to 28nm for the time being and skip to FinFETs next.
Well, there's only so much efficiency and computation magic you can get out at the same node, so I guess we are pretty stuck until that shrink then :( Maybe next year?
It isn't certain yet. In fact recent rumors have pointed to 20nm for Fiji XT, and some have argued that TSMC's 20nm process, despite being optimized for low power, is still perfectly capable of achieving AMD's clockspeed targets. We'll see, but I would be surprised if Fiji XT is even feasible on 28nm if the rumored specs are true.
I'm guessing the yield for 20nm is not too insane for large chips, after all Apple's chip is like 3B transistors. Would be cool if AMD does use 20nm. At least keep things fresh and interesting.
Most rumours state the opposite that R9 300 series will be exclusively on 28nm. 20nm at TSMC is for SoCs and is unlikely to be suitable for flagship performance GPUs.
With 12 GB VRAM in the top model the 384 bit bus is practically certain. From this they can go to 6 GB at full performance. Anything else will be slower after a certain amount of RAM is used up. So you probably rather want a 6 GB GTX980Ti at full performance than an 8 GB version with a handicap.
I think many of you have not understood that it is still better having 8GB as 6GB full speed +2GB slow speed than simply having 6GB full speed. Also the GTX970 may have 0.5GB slow memory partition, but this is still faster than system memory, so it is better to have it than not. People believing that the card would be better with 3.5GB or VRAM simply do not understand how things work.
This was the behavior I inferred from reading the article as well, with the additional constraint that even if you have a high memory bandwidth platform (like Haswell-E), access to main memory is still bottlenecked by a PCI-E 3.0 x16 bus to significantly less than the 0.5GB "slow" partition. I'm assuming that there are significant latency costs associated with reaching out to main memory compared to the local "slow" partition as well, although those are harder to quantify.
It is correct, it is slower than the bulk of the Vram, but still much faster than accessing anything anywhere else in the system.
Those that bitch about the performance can just go in a hole and not come out, because it was a great card for the $$$ when it came out, and guess what? It still is!
Those that are upset about the failure to properly advertise it at all, and instead the insistance that it has the "Same memory structure as the GTX 980" have a valid claim, even if it really is just a chance to try and get something for free, I would jump on too! :P
Except, that's not what happens at all. Either the assets requested are in the VRAM, and they're pointed to and accessed via the memory controller on the GPU, or they're not in the VRAM, and they need to be moved in taking the place of something else. There is absolutely no way in which having that extra RAM is going to cause more stuttering than if it were simply not there.
People's logic amazes me sometimes. Thats like saying you can have 6x 512gb Samsung 850 PRO ssd's AND 2x 850 EVO's (that are slower), and you'd rather have JUST the PRO's.
There was an interesting article published about this recently, but it basically said that the 390X would be limited to 4GB at launch due to the HBM memory config. That's the current limit given a 4096-bit interface and 4-hi 1GB modules. The question is when will Hynix launch stacked modules with 2GB capacities? That's when you'll most likely see an 8GB 390X option.
Perhaps someone who knows more about memory than i do can help me with this. But isn't using the HBM with such a high bandwidth kind of overkill at this point? I.e. it would be like having a water line that can carry 600 gallons/min, and having a pump that can do 550 gallons/min, but then upgrading that pump to one that can do 900 gallons/min.
I mean, right now the memory already isn't the main bottleneck from what i understand, so while you will get performance gains from the HBM, isnt it relatively miniscule?
I just have the feeling its being put in more for a "big number" on the tech specs sheet for the marketing department to use, than for actually significantly improving the video card.
looks nice, hopefully the full vapor chamber for the cooler at that price! "and it will once again be using NVIDIA’s excellent metal cooler and shroud, originally introduced on the original GTX Titan." would read better as "and it will once again be using NVIDIA’s excellent metal cooler and shroud, first introduced on the original GTX Titan."
This thing is going to be beastly. I buy AMD because I don't need the best, and I would like some competition. But Nvidia has just been better in terms of performance for power.
I buy AMD because I want best for the buck and it's hard to find a mid performance GPU (at least in Germany) which isn't shamelessly overpriced on nVidia's end.
Also, if one is going for "the best" in absolute performance terms, power consumption of Radeon 295 shouldn't matter.
Looks good so far. Can't wait to see AMDs new line and their answer to what Nvidia has been throwing out. Hopefully this will come out at a better price point than before but doubt it. So far the Titans have come out at $1000 and $3000 for the dual GPU one.
I think the best we can hope for (and really what would make the most sense) is a replacement at the same $1000 price point. But I have this bad feeling it's going to be significantly more than that. If Nvidia does decide to jack up the price it really makes you appreciate the stability of Intel's SKU's and pricing over the past ~9 years or so (both at the high and low-end), despite the lack of competition above $200-$300.
If the rumors hold that this will be gimped in FP64 performance, I would struggle to see how they could justify the price of $1k. That was one of the reasons, if not the singular reason to charge such a premium. I'm going to swag it at $850.
Not exactly gimped, but if the rumors hold true GM200 won't offer any significant double precision performance boost over GK110. It's rumored that it'll be limited to 1/4 FP32, as opposed to 1/3 on GK110.
If that's the case, then I'll retract my swag. Though I can't say I've run across anything that suggests it will hit 1/4, but a lot of things that suggest Maxwell as a whole is down on FP64 compared to Kepler. That said, if that's truely a limit, then this is not what the Titan was/is/should be, and as a pure gaming card, no matter how great, cannot command that price.
80000000000/x = 26600000 (roughly due to higher density with Maxwell like shown with GM204) 80000000000 = 26600000x x = 80000000000/26600000 GM200 Transistor Count: 3007 Nearest SMM: 23, Rounding up to 24
GM200: 24SMM, 3072 Cores. 384bit, 12GB VRAM Thats my estimation
While your conclusion of 24 SMMs is likely* correct, the assumptions made to get to that figure are wrong. The core counts you've listed are FP32 cores. There's a WHOLE lot more going on in a GPU that FP32 cores. ROPs, memory controllers, display transmitters, and more are part of the transistor count. GK1xx and GM2xx don't even share the same ratio of FP32 to FP64 cores. Calculating the number of transistors per core from the total figure just isn't right.
detecting a small patern lol but i dont think the shaders per SMM are going to stay the same or the size, the Gm 206 is a rectangle because it only has one row of SMMs the Gm 204 is a square because it has 2 rows the Gm 200 would be too tall, and not a square if it had 3 rows, wich it would make it unoptimal, and maybe even impossible to manufacture given the size. A square die is almost certain, lets see how they do it.
The number of FP32 cores per SMM in GM200 is almost certainly the same as in GM204. What's up in the air is the number of FP64 cores per SMM. I can't imagine they'd stick with the paltry figure in GM204 and call this thing a Titan, so expect changes there.
There's also nothing that says that repeated blocks have to be arranged in a fixed grid. It's just convenient to do so. Look at die shots of GK110. They can pretty much do whatever they damn well please...within reason.
in the GK110 case 3 rows tall was just about perfect for the square die, in this case 2 is too litle and 3 is too much, unless they lay down one row, they ar going to adjust the SMMs tall/wide relation
Why 3 x 8 when 4 x 6 is an option? Also, they could do a 5 x 5 grid with the 25th space taken up by ROPs, GPC common hardware, crossbar hardware, whatever else they want.
The GAMERS who bought the original Titan should have done their own research and realized that was not a product for them, just like all successive Titans. The regret they felt was all on them.
I should also mention that I believe gamers will buy whatever they want. Just realize that the product you're paying extra cash for has a different reason for its price than gaming. It's a heavier compute card that just so happens to also do gaming pretty well. Gaming alone on a Titan is a waste of the extra resources provided for workloads that gaming won't utilize.
But no one is whining it's not future proof because of lack of memory.
Yes, I know, the core can't do it anyway (future games high res with massive mem useage and decent frames) , but telling the lemmings that doesn't work.
You write in the article Ryan Smith "and it will once again be using NVIDIA’s excellent metal cooler and shroud, >originally< introduced on the original GTX Titan." This is nonsense, the GTX 690 had it 4 months before TITAN.
The design for the 690 cooler was adjusted for the original Titan. The concepts existed 4 months before the Titan, but that specific cooler was created for the Titan.
Funny you should say that. Just last week the German magazin "PC Games Hardware" was showing off some UHD Gaming Benchmarks with a 8GB FirePro and a 4GB 290X, and they noticed the FirePro using more than 6GB of VRAM, while the 290X was using an additional 1.5GB of System Memory.
And we're just getting started to game in UHD, I think 12GB will be very much appreciated by enthusiast gamers over the next 2 years.
At 4K resolution, Ultra settings plus 4xAA, BF4 uses 3.167GB VRAM, while Crysis 3 uses 3.668GB.
The main reason for the recent jump in VRAM usage is not 4K, but the fact that the new gen consoles have a uniform memory architecture to compensate for their otherwise much inferior hardware, and ports to PC are done lazily as always.
That is why, 8GB VRAM is going to be the gold standard for the top-tier GPUs in the coming years. Putting an extra 4GB is just going to be wasteful, 4K or not, and the end users would be better off with higher bandwidth/lower latency 8GB. On the other hand, I suppose Nvidia is going to have to justify the ~$1,500 price tag, where that extra 4GB would be helpful.
It all depends on what these new APIs DX12/Vulkan deliver, DX12 stacks VRAM so that should extend the usefulness of most DX12 capable GPUs, anyone with a Fermi, Kepler, and full DX12 Maxwell will be very happy their GPU investment in SLI. One big reason prices have come down is that TSMC's 28nm process is mature offering high yields, I expect Titan X prices to fall sharply after a few months that's when I'll time my new build, Skylake and Titan X will bring great performance for many years to come and I will add another Titan X for a SLI upgrade boost when needed at even lower prices. All GPUs when they are first released have a high price tag but drop within a few months, its all good for the patient gamer.
Enthusiasts who can afford $1K cards in pairs don't keep them for 2 years sorry. By the time 12GB on a card is useful, the Titan X's performance could be purchased in a $300 Pascal/Volta. In 12 months alone we went from $700 GTX780Ti to $330 970 and $1500 R9 295X2 to $600 today. Anyone who buys $2000-2500 worth of graphics cards to "future proof" is doing it wrong. Most people who will buy this are those for whom $1350 or so is not a lot of money, they use it for work/PhD research/universities or early adopters who flip cards before they tank in value too much, which means selling it well before the 2 year mark.
Anandtech, please, I beg of you, use the internationally recognised ISO 8601 yyyy-mm-dd date format. Your current date format is confusing for non-Americans to read.
Anandtech is based in the US so of course it's date format will be the US format. If we have to suffer the crappy format on euro websites you will suffer on US websites
I'm sure you can figure out that it's not yet April 3rd 2015. As an American in the UK, it's not too hard to figure out the context of the date based on what site I'm looking at.
As far as current posting dates go (not historical,) you're right Railgun. I was referring to the launch dates. Not these two specific launch dates, obviously, but my Date Format Frustration Level just happened to be reached with this specific article.
It stinks about being stuck at 28nm for so long. Despite that, the Titans have always been pretty impressive hardware but at 10x the price of an inexpensive Windows tablet like the TW802 or Stream 7, I think my next computer will have 4 Intel Gen7 graphics execution units in it and I'll just stick to managing my expectations for entertainment and wait until passively cooled budget processors contain faster graphics in a few years. It just seems so hard to rationalize a single component purchase that gets installed inside a non-portable box that's not really any better at being a word processor than a phone or tablet I can stick in a purse. They'd both run Terraria and let me crush the same candy.
Would you like to drive a [whatever brand of small cheap scooter/car is looked down upon in your circle of friends] or would you prefer a Tesla/Ferrari/Range Rover? They both get you from point A to point B.
Your analogy could use a little tweak. Big GPUs are doing more work than small GPUs. It'd be closer to say when you're going on vacation, do you bring a Smart car with two people and a bag or bring a Crossover/SUV with skis, tents, and whatever else you want. Bring all of the things.
I don't let the opinions of my friends influence what I buy with the money I earn. :) I buy the stuff that I wanna which I think will fill my needs. And yeah, actually I would totally prefer an inexpensive car over some huge box or sports car since I also don't measure my status with the objects I own. It's lots more important, I think, to have lots of liquid money that can be used to generate compounding interest rather than loans or debt on things that might keep up with inflation (if they're real estate) or will just progressively lose value over time like a car of all things.
This comparison is rather flawed. But let's talk from a performance point of view. In 12 months from now a $600 card will be as fast as this Titan X and in 12 more months a $350-400 card. In 36 more months you'll be able to buy a card with similar performance for $200. If you go out and buy a McLaren 650S or a Ferrari 488 GTB, can you buy a brand new 2018 McLaren/Ferrari with similar performance in 36 months for 5-7X less? No. Many people at AT could buy a $1350 card but they don't do so because of how quickly GPUs drop in price/are superseded. Look at the original $1K Titan. 2 years from its launch a $240 R9 290 and a $310 970 both beat it. In 9 months from the original Titan launch a $400 R9 290 matched it. That's why most gamers won't buy $2000-2500 worth of Titan Xs without waiting for R9 390 series/GM200 consumer 6GB card.
These price comparisons are flawed, however. If I'm in the market for a video card right now, I can buy a 600 top of the line card, or a 300 dollar very good card. Then, in a year, the 300 dollar cards perform as well or slightly better than the 600 card of today. But the 300 cards today will not age as well. So in a year, to maintain reasonable performance, I'll be in the market for another card. So I have to decide then to buy either a 300 dollar card then, or a more expensive one. Either way, after two years, I'll have comparable performance whether I spent 300 twice in two years, or 600 once. The difference is that I have better performance for the first year. I know that prices don't always work out that way, but making a blanket statement that buying an expensive product now is a bad idea since similar performance is available for half the cost in a year is ultimately flawed. The 500 card I bought in 2013 is still doing very well for me, enough so that the current 250 cards aren't really any better, at least performance wise. And I got solid performance for a full year.
What do you expect to happen in you're liffetimes ? Fying cars, fusion energy. Be greatful you have the choice not to buy this and wait for something better before you die. Sorry to be morbid but i come from 1982 and nintendo and a 48666dx2 gateway. Dont hate and shit on everything and enjoy the gifts you all have availablbe. Love anantec reader since early 2000s
486 processors weren't available for purchase until the very late 1980s...about 1989, but they weren't ticking along at 66MHz until the early 90s...about 1993 so I think 1982 might be a bit off the mark as far as the availability that particular CPU is concerned. On a side note, my first computer was a Commodore 64. ;)
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
104 Comments
Back to Article
inighthawki - Wednesday, March 4, 2015 - link
The table incorrectly lists "GTX Titan Black" twice. I presume the first column should read "GTX Titan X"?blanarahul - Wednesday, March 4, 2015 - link
Big Maxwell on 28nm?! Whooo! Someone's desperate!bunnyfubbles - Wednesday, March 4, 2015 - link
desperate for the fabs to get on the ball and beyond 28nm. And while 28nm is likely, its a good best guess, not yet confirmed, although Titan X debuting on something other than 28nm would likely be bigger news than the chip itself.FlushedBubblyJock - Wednesday, March 11, 2015 - link
The GTX980 Maxwell has 5.2B transistors, while the Maxwell TX has 8.0B transistors, so that's quite a leap.Flunk - Wednesday, March 4, 2015 - link
They have to, TSMC isn't doing a high-power 20nm-class process. They're skipping that node.blanarahul - Wednesday, March 4, 2015 - link
WHAT?! WHEN DID THAT HAPPEN?! So what will 3rd (or 4th) gen Maxwell be based on? 16FF or 16FF+?blanarahul - Wednesday, March 4, 2015 - link
Just googled. 16/14nm FF in 2016. Sigh. Well atleast we'll get to see how efficient can we really be on 28nm.nandnandnand - Wednesday, March 4, 2015 - link
28nm: the beatingist, deadest horsedanjw - Friday, March 6, 2015 - link
It has been talked about since last year. TSMC doesn't seem to be able to get higher powered processors on its 20nm process. It isn't clear if this was a decision on their part or that they found it just didn't work.dragonsqrrl - Wednesday, March 4, 2015 - link
This has been known for about half a year now. GM200 is being fabbed on 28nm. What hasn't been known is the exact transistor count and die size. Rumors suggest it'll be the largest GPU Nvidia's produced, ~600mm².blanarahul - Wednesday, March 4, 2015 - link
That's even larger than GT200 (576 mm^2). My calculations put it around ~610 mm^2 which is gigantic.Kevin G - Wednesday, March 4, 2015 - link
My guestimation put it at 594 mm^2, pretty close. Remember that not all functional units are being increased (video decoder/encoder etc.).Still not a record holder for chip size. Intel's Tukwila Itanium 2 was 699 mm^2. At the time, the 65 nm limit was estimated around 750 mm^2 due to the optics involved.
eanazag - Wednesday, March 4, 2015 - link
We need to ask if the VRAM is partitioned to two 6 GB sets or three 4 GB sets or one 12 GB. Fair question. TDP and price are the only items from the table above that I'm curious about. The each Titan chip has spent too little time as the top dog for the money they're asking. It maybe makes sense as a CUDA card, but it has none of the virtualization features.It is total crap that 8 generations of GPUs need to be 28 nm. I totally understand it may not be AMD or Nvidia's fault. Too many processes are built for low power chips.
To be mentioned - the AMD R295 is now about $600 on Newegg. Don't forget it will cost $30 a month in electrical bills though.
nevcairiel - Wednesday, March 4, 2015 - link
If the number of ROPs is evenly divisible with the amount of memory, there doesn't need to be any partitioning. Its only a tool used when the numbers diverge, like 56 ROPs and 4GB.However 96 ROPs on a 384-bit bus match perfectly with 6 or 12 GB of memory. Extrapolating from the 980, it will have 12 ports with a 32-bit MC each and 8 ROPs, each hosting 128MB of memory.
Kutark - Sunday, March 8, 2015 - link
I read somewhere else it was going to be a contiguous 12gb, but so far noone knows for certain.It would be a little ridiculous for them to put that much RAM on it and not be able to properly utilize it though.
Refuge - Monday, March 9, 2015 - link
lol this isn't the 970...Kutark - Sunday, March 8, 2015 - link
Desperate for what? Gaining an even larger lead on AMD?Wreckage - Wednesday, March 4, 2015 - link
I'm glad someone had new hardware to show at GDC. Now if someone could mail me 2 of these for "testing purposes".ArmedandDangerous - Wednesday, March 4, 2015 - link
First column should say Titan X and not Titan Black and then Titan Black again :Dsquatsh - Wednesday, March 4, 2015 - link
Okay so GM204 (980) has 64 rops. This has 96 (with a question mark next to it ) so thats 150%Adding a extra 50% to the other numbers from GM204 gives us:
3077 Cuda Cores
192 Texture Units
384 bit bus is 150% of 256% of the 980
This has 3 times the RAM, but it is a titan, so maybe 980ti will have 6gb (ie 150% of 980 again)
Memory clock and core clock I would expect to be the usual. 7ghz and 1ghz-ish respectively.
Oh and the 5.2 Billion transistors of the 980 * 150% would give 7.8, or close to 8 billion. So this times by 150% thing seem to fit pretty well.
(This is all from just noticing the simple 50% increase in ROPs and doing the numbers incase someone is interested, none of this is solid information).
squatsh - Wednesday, March 4, 2015 - link
3072 Cuda cores*smartypnt4 - Wednesday, March 4, 2015 - link
I completely agree. These are the specs the leaks have pointed to as well.I'll also note this: this will be a 24-SMM design, with all SMMs on the die enabled, as 7.8B is around the upper limit of what TSMC can put on a single die right now at 28nm.
This means there won't be a new version of this chip like there was with the original Titan and Titan Black.
I also expect there to be a consumer version of this with 22 SMM's enabled, with the requisite ROPs and L2 cache disabled like they were on the 970 It'll likely also have 6GB of VRAM (For yes, 5GB of fast VRAM or however the math works out for their bus partitioning). This Titan Lite or 980 Ti or 985 or 990 or whatever they're gonna call it will be the real card to get for 4K from NVIDIA until they go through a process shrink. Two of those will be quite potent.
will54 - Wednesday, March 4, 2015 - link
actually if its a 384 bit bus it should be a full 6 GB'ssmartypnt4 - Wednesday, March 4, 2015 - link
The PCB will have 6GB, but it could be partitioned similarly to the 970's, which has a 3.5GB high-performance partition and a slow 512MB partition. I haven't worked out exactly what the equivalent would be based on SMM configuration, but it wouldn't surprise me to see 6GB on the PCB itself with something like a 768MB slow memory partition and 5.25GB fast memory partition.Or they'll do something totally new. No idea really. It's hard to even speculate what they might do. For the 22 SMM, it's conceivable they'll partition the memory like on the 970, but for the 20 SMM version, they could just disable one of the L2/ROP/Memory Controller units and go with a 320-bit bus for that part with 5GB of VRAM. That would actually be a really good part to have between the 980 and the other GM200-based cards.
Yojimbo - Wednesday, March 4, 2015 - link
This is a "prosumer" card which is sold with the idea that it could be used for things other than games, hence the FP64 at 1/3 FP32 for the other two Titan cards. It's very unlikely they would do such memory partitioning since, although 12GB is not going to be used for gaming, it might be used in other situations.squatsh - Wednesday, March 4, 2015 - link
The strange memory speeds in the 970 are due to the ROP set up now the SMMs, which can be completely independent. So you can have 22 SMMs and full speed memory. Overall the whole memorygate or whatever it is called was a bit of a mess, but as a technology the extra granularity is good for Nvidia (for yields) and us (they can be less conservative,ie 4gb of 3.5fast and .5 slo instead of just plain old 3gb). But I agree that while they are chopping off 2 SMMs for yields, they might mess around with the memory system again.squatsh - Wednesday, March 4, 2015 - link
(chopping off memory bits for the consumer lite version, if there will be one).Flunk - Wednesday, March 4, 2015 - link
I've been hoping for a 8GB 980TI and this leaves them room for that. They're going to have to do that if AMD brings out a 8GB 390x, which they probably will.hammer256 - Wednesday, March 4, 2015 - link
So is AMD going to use 20nm node for their GPUs? Or are they like Nvidia and sticking to the 28 until 16/14nm FFET?nevcairiel - Wednesday, March 4, 2015 - link
The 20nm process isn't good enough for high performance GPUs, so the consensus is that there won't be 20nm dGPUs, but they'll stick to 28nm for the time being and skip to FinFETs next.hammer256 - Wednesday, March 4, 2015 - link
Well, there's only so much efficiency and computation magic you can get out at the same node, so I guess we are pretty stuck until that shrink then :( Maybe next year?dragonsqrrl - Wednesday, March 4, 2015 - link
It isn't certain yet. In fact recent rumors have pointed to 20nm for Fiji XT, and some have argued that TSMC's 20nm process, despite being optimized for low power, is still perfectly capable of achieving AMD's clockspeed targets. We'll see, but I would be surprised if Fiji XT is even feasible on 28nm if the rumored specs are true.MrSpadge - Wednesday, March 4, 2015 - link
Hitting AMDs frequency targets is not so hard when you consider they're running Tonga on R9 285 at 925 MHz. And this is not even using low voltage.hammer256 - Wednesday, March 4, 2015 - link
I'm guessing the yield for 20nm is not too insane for large chips, after all Apple's chip is like 3B transistors. Would be cool if AMD does use 20nm. At least keep things fresh and interesting.RussianSensation - Friday, March 6, 2015 - link
Most rumours state the opposite that R9 300 series will be exclusively on 28nm. 20nm at TSMC is for SoCs and is unlikely to be suitable for flagship performance GPUs.MrSpadge - Wednesday, March 4, 2015 - link
With 12 GB VRAM in the top model the 384 bit bus is practically certain. From this they can go to 6 GB at full performance. Anything else will be slower after a certain amount of RAM is used up. So you probably rather want a 6 GB GTX980Ti at full performance than an 8 GB version with a handicap.CiccioB - Wednesday, March 4, 2015 - link
I think many of you have not understood that it is still better having 8GB as 6GB full speed +2GB slow speed than simply having 6GB full speed.Also the GTX970 may have 0.5GB slow memory partition, but this is still faster than system memory, so it is better to have it than not. People believing that the card would be better with 3.5GB or VRAM simply do not understand how things work.
MrCommunistGen - Wednesday, March 4, 2015 - link
This was the behavior I inferred from reading the article as well, with the additional constraint that even if you have a high memory bandwidth platform (like Haswell-E), access to main memory is still bottlenecked by a PCI-E 3.0 x16 bus to significantly less than the 0.5GB "slow" partition. I'm assuming that there are significant latency costs associated with reaching out to main memory compared to the local "slow" partition as well, although those are harder to quantify.Refuge - Monday, March 9, 2015 - link
It is correct, it is slower than the bulk of the Vram, but still much faster than accessing anything anywhere else in the system.Those that bitch about the performance can just go in a hole and not come out, because it was a great card for the $$$ when it came out, and guess what? It still is!
Those that are upset about the failure to properly advertise it at all, and instead the insistance that it has the "Same memory structure as the GTX 980" have a valid claim, even if it really is just a chance to try and get something for free, I would jump on too! :P
Oxford Guy - Wednesday, March 4, 2015 - link
Right, because it's advantageous to have XOR contention and 28 gb/s bandwidth that causes microstutter and slows the performance of the fast VRAM.PatHeist - Friday, March 6, 2015 - link
Except, that's not what happens at all. Either the assets requested are in the VRAM, and they're pointed to and accessed via the memory controller on the GPU, or they're not in the VRAM, and they need to be moved in taking the place of something else. There is absolutely no way in which having that extra RAM is going to cause more stuttering than if it were simply not there.Kutark - Sunday, March 8, 2015 - link
People's logic amazes me sometimes. Thats like saying you can have 6x 512gb Samsung 850 PRO ssd's AND 2x 850 EVO's (that are slower), and you'd rather have JUST the PRO's.dragonsqrrl - Wednesday, March 4, 2015 - link
There was an interesting article published about this recently, but it basically said that the 390X would be limited to 4GB at launch due to the HBM memory config. That's the current limit given a 4096-bit interface and 4-hi 1GB modules. The question is when will Hynix launch stacked modules with 2GB capacities? That's when you'll most likely see an 8GB 390X option.Laststop311 - Wednesday, March 4, 2015 - link
HBM 1st gen is limited to 4GB so 390x will only be 4GBKutark - Sunday, March 8, 2015 - link
Perhaps someone who knows more about memory than i do can help me with this. But isn't using the HBM with such a high bandwidth kind of overkill at this point? I.e. it would be like having a water line that can carry 600 gallons/min, and having a pump that can do 550 gallons/min, but then upgrading that pump to one that can do 900 gallons/min.I mean, right now the memory already isn't the main bottleneck from what i understand, so while you will get performance gains from the HBM, isnt it relatively miniscule?
I just have the feeling its being put in more for a "big number" on the tech specs sheet for the marketing department to use, than for actually significantly improving the video card.
FlushedBubblyJock - Wednesday, March 11, 2015 - link
It should be faster accessing it, so it's better - at least that's the plan.dstockwell23 - Wednesday, March 4, 2015 - link
looks nice, hopefully the full vapor chamber for the cooler at that price!"and it will once again be using NVIDIA’s excellent metal cooler and shroud, originally introduced on the original GTX Titan."
would read better as
"and it will once again be using NVIDIA’s excellent metal cooler and shroud, first introduced on the original GTX Titan."
ingwe - Wednesday, March 4, 2015 - link
I love the launch price!This thing is going to be beastly. I buy AMD because I don't need the best, and I would like some competition. But Nvidia has just been better in terms of performance for power.
Yojimbo - Wednesday, March 4, 2015 - link
Buying something you don't want for competition is sort of like having no competition to begin with.sr1030nx - Wednesday, March 4, 2015 - link
Agreed, launch price is pretty funny.medi03 - Thursday, March 5, 2015 - link
I buy AMD because I want best for the buck and it's hard to find a mid performance GPU (at least in Germany) which isn't shamelessly overpriced on nVidia's end.Also, if one is going for "the best" in absolute performance terms, power consumption of Radeon 295 shouldn't matter.
Yojimbo - Thursday, March 5, 2015 - link
Who says "the best" means the best in absolute performance terms? His message seemed to suggest he didn't consider it that way.r3loaded - Wednesday, March 4, 2015 - link
Inb4 11.5GB of RAM.dragonsqrrl - Wednesday, March 4, 2015 - link
Inb4 the next parrot who doesn't understand Maxwell's ROP/memory system.Crunchy005 - Wednesday, March 4, 2015 - link
Looks good so far. Can't wait to see AMDs new line and their answer to what Nvidia has been throwing out. Hopefully this will come out at a better price point than before but doubt it. So far the Titans have come out at $1000 and $3000 for the dual GPU one.hammer256 - Wednesday, March 4, 2015 - link
I would imagine $1k is probably the reasonable expectation here :(dragonsqrrl - Wednesday, March 4, 2015 - link
I think the best we can hope for (and really what would make the most sense) is a replacement at the same $1000 price point. But I have this bad feeling it's going to be significantly more than that. If Nvidia does decide to jack up the price it really makes you appreciate the stability of Intel's SKU's and pricing over the past ~9 years or so (both at the high and low-end), despite the lack of competition above $200-$300.CaedenV - Wednesday, March 4, 2015 - link
Launch Price: A large numberLMAO, that made my day!
Railgun - Wednesday, March 4, 2015 - link
If the rumors hold that this will be gimped in FP64 performance, I would struggle to see how they could justify the price of $1k. That was one of the reasons, if not the singular reason to charge such a premium. I'm going to swag it at $850.dragonsqrrl - Wednesday, March 4, 2015 - link
Not exactly gimped, but if the rumors hold true GM200 won't offer any significant double precision performance boost over GK110. It's rumored that it'll be limited to 1/4 FP32, as opposed to 1/3 on GK110.Railgun - Thursday, March 5, 2015 - link
If that's the case, then I'll retract my swag. Though I can't say I've run across anything that suggests it will hit 1/4, but a lot of things that suggest Maxwell as a whole is down on FP64 compared to Kepler. That said, if that's truely a limit, then this is not what the Titan was/is/should be, and as a pure gaming card, no matter how great, cannot command that price.huaxshin - Wednesday, March 4, 2015 - link
GK104: 3.54billion Transistors. 1536 cores. 23046875/coreGM204: 5.2billion Transistors. 2048 cores. 25390625/core
GK110: 7.1billion Transistors. 2880 cores. 24652777/core
GM200: 8.0billion Transistors. x cores
80000000000/x = 26600000 (roughly due to higher density with Maxwell like shown with GM204)
80000000000 = 26600000x
x = 80000000000/26600000
GM200 Transistor Count: 3007
Nearest SMM: 23, Rounding up to 24
GM200: 24SMM, 3072 Cores. 384bit, 12GB VRAM
Thats my estimation
Urizane - Wednesday, March 4, 2015 - link
While your conclusion of 24 SMMs is likely* correct, the assumptions made to get to that figure are wrong. The core counts you've listed are FP32 cores. There's a WHOLE lot more going on in a GPU that FP32 cores. ROPs, memory controllers, display transmitters, and more are part of the transistor count. GK1xx and GM2xx don't even share the same ratio of FP32 to FP64 cores. Calculating the number of transistors per core from the total figure just isn't right.likely* = So likely that it's almost a certainty.
huaxshin - Wednesday, March 4, 2015 - link
GK110 is there ;)nunomoreira10 - Wednesday, March 4, 2015 - link
Gm 206: 1024 shadersGm 204: 2048 shaders
Gm 200: 3072 shaders
detecting a small patern lol
but i dont think the shaders per SMM are going to stay the same or the size,
the Gm 206 is a rectangle because it only has one row of SMMs
the Gm 204 is a square because it has 2 rows
the Gm 200 would be too tall, and not a square if it had 3 rows, wich it would make it unoptimal, and maybe even impossible to manufacture given the size.
A square die is almost certain, lets see how they do it.
Urizane - Wednesday, March 4, 2015 - link
The number of FP32 cores per SMM in GM200 is almost certainly the same as in GM204. What's up in the air is the number of FP64 cores per SMM. I can't imagine they'd stick with the paltry figure in GM204 and call this thing a Titan, so expect changes there.There's also nothing that says that repeated blocks have to be arranged in a fixed grid. It's just convenient to do so. Look at die shots of GK110. They can pretty much do whatever they damn well please...within reason.
nunomoreira10 - Wednesday, March 4, 2015 - link
in the GK110 case 3 rows tall was just about perfect for the square die, in this case 2 is too litle and 3 is too much, unless they lay down one row, they ar going to adjust the SMMs tall/wide relationUrizane - Thursday, March 5, 2015 - link
Why 3 x 8 when 4 x 6 is an option? Also, they could do a 5 x 5 grid with the 25th space taken up by ROPs, GPC common hardware, crossbar hardware, whatever else they want.TheRealAnalogkid - Wednesday, March 4, 2015 - link
IIRC, wasn't the GTX 690 the first with the metal cooler and shroud, albeit central fan location?I'm betting this will be $1500- and plenty of game applications with 4k monitors/60+ refresh.
deeps6x - Wednesday, March 4, 2015 - link
Gonna be another $999 marketing gimmick that nobody actually buys, or regrets 2 months later if they do, when the GTX equivalent comes out?That said, good on ya Nvidia. Keep forcing AMD to compete.
Urizane - Thursday, March 5, 2015 - link
The GAMERS who bought the original Titan should have done their own research and realized that was not a product for them, just like all successive Titans. The regret they felt was all on them.Urizane - Thursday, March 5, 2015 - link
I should also mention that I believe gamers will buy whatever they want. Just realize that the product you're paying extra cash for has a different reason for its price than gaming. It's a heavier compute card that just so happens to also do gaming pretty well. Gaming alone on a Titan is a waste of the extra resources provided for workloads that gaming won't utilize.FlushedBubblyJock - Wednesday, March 11, 2015 - link
But no one is whining it's not future proof because of lack of memory.Yes, I know, the core can't do it anyway (future games high res with massive mem useage and decent frames) , but telling the lemmings that doesn't work.
deeps6x - Wednesday, March 4, 2015 - link
Ya'all really should post a 980 and it's specs on that chart as well for reference purposes.Innokentij - Wednesday, March 4, 2015 - link
You write in the article Ryan Smith "and it will once again be using NVIDIA’s excellent metal cooler and shroud, >originally< introduced on the original GTX Titan." This is nonsense, the GTX 690 had it 4 months before TITAN.Urizane - Thursday, March 5, 2015 - link
The design for the 690 cooler was adjusted for the original Titan. The concepts existed 4 months before the Titan, but that specific cooler was created for the Titan.Urizane - Thursday, March 5, 2015 - link
How did my comment get here? I meant to reply to the 690 cooler comment.Urizane - Thursday, March 5, 2015 - link
OMG. The live comment updates showed my comment being in reply to "You a word." There's a bug, but good luck trying to replicate it.abianand - Thursday, March 5, 2015 - link
damn, when i looked at the launch price of the Titan X i almost the coffee i was drinking on my keyboard.kyuu - Thursday, March 5, 2015 - link
You a word.abianand - Thursday, March 5, 2015 - link
@kyuu: lol, thx. took some time to understand.wish thr was an edit button
D. Lister - Thursday, March 5, 2015 - link
Latching 12GB of VRAM on a gaming GPU is just a terrible waste of resources. Keep the extra 4GBs Nvidia, and give us more raw horsepower instead.ShieTar - Thursday, March 5, 2015 - link
Funny you should say that. Just last week the German magazin "PC Games Hardware" was showing off some UHD Gaming Benchmarks with a 8GB FirePro and a 4GB 290X, and they noticed the FirePro using more than 6GB of VRAM, while the 290X was using an additional 1.5GB of System Memory.And we're just getting started to game in UHD, I think 12GB will be very much appreciated by enthusiast gamers over the next 2 years.
D. Lister - Thursday, March 5, 2015 - link
http://www.digitalstormonline.com/unlocked/video-m...At 4K resolution, Ultra settings plus 4xAA, BF4 uses 3.167GB VRAM, while Crysis 3 uses 3.668GB.
The main reason for the recent jump in VRAM usage is not 4K, but the fact that the new gen consoles have a uniform memory architecture to compensate for their otherwise much inferior hardware, and ports to PC are done lazily as always.
That is why, 8GB VRAM is going to be the gold standard for the top-tier GPUs in the coming years. Putting an extra 4GB is just going to be wasteful, 4K or not, and the end users would be better off with higher bandwidth/lower latency 8GB. On the other hand, I suppose Nvidia is going to have to justify the ~$1,500 price tag, where that extra 4GB would be helpful.
CPUGPUGURU - Saturday, March 7, 2015 - link
It all depends on what these new APIs DX12/Vulkan deliver, DX12 stacks VRAM so that should extend the usefulness of most DX12 capable GPUs, anyone with a Fermi, Kepler, and full DX12 Maxwell will be very happy their GPU investment in SLI. One big reason prices have come down is that TSMC's 28nm process is mature offering high yields, I expect Titan X prices to fall sharply after a few months that's when I'll time my new build, Skylake and Titan X will bring great performance for many years to come and I will add another Titan X for a SLI upgrade boost when needed at even lower prices. All GPUs when they are first released have a high price tag but drop within a few months, its all good for the patient gamer.RussianSensation - Friday, March 6, 2015 - link
Enthusiasts who can afford $1K cards in pairs don't keep them for 2 years sorry. By the time 12GB on a card is useful, the Titan X's performance could be purchased in a $300 Pascal/Volta. In 12 months alone we went from $700 GTX780Ti to $330 970 and $1500 R9 295X2 to $600 today. Anyone who buys $2000-2500 worth of graphics cards to "future proof" is doing it wrong. Most people who will buy this are those for whom $1350 or so is not a lot of money, they use it for work/PhD research/universities or early adopters who flip cards before they tank in value too much, which means selling it well before the 2 year mark.AndrewJacksonZA - Thursday, March 5, 2015 - link
Anandtech, please, I beg of you, use the internationally recognised ISO 8601 yyyy-mm-dd date format. Your current date format is confusing for non-Americans to read.*PRETTY* please?
https://en.wikipedia.org/wiki/ISO_8601
Laststop311 - Thursday, March 5, 2015 - link
Anandtech is based in the US so of course it's date format will be the US format. If we have to suffer the crappy format on euro websites you will suffer on US websitesRailgun - Thursday, March 5, 2015 - link
I'm sure you can figure out that it's not yet April 3rd 2015. As an American in the UK, it's not too hard to figure out the context of the date based on what site I'm looking at.AndrewJacksonZA - Thursday, March 5, 2015 - link
As far as current posting dates go (not historical,) you're right Railgun. I was referring to the launch dates. Not these two specific launch dates, obviously, but my Date Format Frustration Level just happened to be reached with this specific article.The Radeon 6400 was launched on 02/07/2011. What date is that exactly? ( https://en.wikipedia.org/wiki/Radeon_HD_6000_Serie... )
Railgun - Thursday, March 5, 2015 - link
Between Feb and July 2011. ;)AndrewJacksonZA - Friday, March 6, 2015 - link
:-)blzd - Thursday, March 5, 2015 - link
12 Gigs of RAM so you know it'll be fast! /sI've been waiting for 12GB of VRAM before making my next GPU purchase, should be a nice jump from my 1GB. /s
BrokenCrayons - Thursday, March 5, 2015 - link
It stinks about being stuck at 28nm for so long. Despite that, the Titans have always been pretty impressive hardware but at 10x the price of an inexpensive Windows tablet like the TW802 or Stream 7, I think my next computer will have 4 Intel Gen7 graphics execution units in it and I'll just stick to managing my expectations for entertainment and wait until passively cooled budget processors contain faster graphics in a few years. It just seems so hard to rationalize a single component purchase that gets installed inside a non-portable box that's not really any better at being a word processor than a phone or tablet I can stick in a purse. They'd both run Terraria and let me crush the same candy.AndrewJacksonZA - Thursday, March 5, 2015 - link
Would you like to drive a [whatever brand of small cheap scooter/car is looked down upon in your circle of friends] or would you prefer a Tesla/Ferrari/Range Rover? They both get you from point A to point B.Urizane - Thursday, March 5, 2015 - link
Your analogy could use a little tweak. Big GPUs are doing more work than small GPUs. It'd be closer to say when you're going on vacation, do you bring a Smart car with two people and a bag or bring a Crossover/SUV with skis, tents, and whatever else you want. Bring all of the things.BrokenCrayons - Friday, March 6, 2015 - link
I don't let the opinions of my friends influence what I buy with the money I earn. :) I buy the stuff that I wanna which I think will fill my needs. And yeah, actually I would totally prefer an inexpensive car over some huge box or sports car since I also don't measure my status with the objects I own. It's lots more important, I think, to have lots of liquid money that can be used to generate compounding interest rather than loans or debt on things that might keep up with inflation (if they're real estate) or will just progressively lose value over time like a car of all things.RussianSensation - Friday, March 6, 2015 - link
This comparison is rather flawed. But let's talk from a performance point of view. In 12 months from now a $600 card will be as fast as this Titan X and in 12 more months a $350-400 card. In 36 more months you'll be able to buy a card with similar performance for $200. If you go out and buy a McLaren 650S or a Ferrari 488 GTB, can you buy a brand new 2018 McLaren/Ferrari with similar performance in 36 months for 5-7X less? No. Many people at AT could buy a $1350 card but they don't do so because of how quickly GPUs drop in price/are superseded. Look at the original $1K Titan. 2 years from its launch a $240 R9 290 and a $310 970 both beat it. In 9 months from the original Titan launch a $400 R9 290 matched it. That's why most gamers won't buy $2000-2500 worth of Titan Xs without waiting for R9 390 series/GM200 consumer 6GB card.erple2 - Sunday, March 8, 2015 - link
These price comparisons are flawed, however. If I'm in the market for a video card right now, I can buy a 600 top of the line card, or a 300 dollar very good card. Then, in a year, the 300 dollar cards perform as well or slightly better than the 600 card of today. But the 300 cards today will not age as well. So in a year, to maintain reasonable performance, I'll be in the market for another card. So I have to decide then to buy either a 300 dollar card then, or a more expensive one. Either way, after two years, I'll have comparable performance whether I spent 300 twice in two years, or 600 once. The difference is that I have better performance for the first year. I know that prices don't always work out that way, but making a blanket statement that buying an expensive product now is a bad idea since similar performance is available for half the cost in a year is ultimately flawed. The 500 card I bought in 2013 is still doing very well for me, enough so that the current 250 cards aren't really any better, at least performance wise. And I got solid performance for a full year.FlushedBubblyJock - Wednesday, March 11, 2015 - link
Thank you for the logic.MyNuts - Monday, March 9, 2015 - link
What do you expect to happen in you're liffetimes ? Fying cars, fusion energy. Be greatful you have the choice not to buy this and wait for something better before you die. Sorry to be morbid but i come from 1982 and nintendo and a 48666dx2 gateway. Dont hate and shit on everything and enjoy the gifts you all have availablbe. Love anantec reader since early 2000sBrokenCrayons - Tuesday, March 10, 2015 - link
486 processors weren't available for purchase until the very late 1980s...about 1989, but they weren't ticking along at 66MHz until the early 90s...about 1993 so I think 1982 might be a bit off the mark as far as the availability that particular CPU is concerned. On a side note, my first computer was a Commodore 64. ;)MyNuts - Monday, March 9, 2015 - link
wow thanks for blocking my comment. great censorship. try not to be late with news pleaseMyNuts - Monday, March 9, 2015 - link
If you dont have the money to get things like this its not a bad thing, just unfortunate. Im fortunate