Titan X Announced

[UaVaj]

The 980 didn't have a memory issue. That was the 970.

The 980 architecture is fine. But it is also a smaller die.

This is the GM200, the full-size die, not the cut-down versions that have populated current Maxwell consumer parts. It's a single-chip card with 12GB of direct memory, there won't be any issues addressing any of it.

please link your source.

no way this is a single gpu card with true 12gb of vram.

if so. that would be significant boost in performance.
in that case. want 4 for 7680x1440.

 

[96Firebird]

The GPU underlying GTX Titan X is 8 billion transistors, which similar to the original GTX Titan’s launch means we’re almost certainly looking at Big Maxwell. NVIDIA will be pairing it with 12GB VRAM – indicating a 384-bit memory bus – and it will once again be using NVIDIA’s excellent metal cooler and shroud, originally introduced on the original GTX Titan.

Source

 

[RaulF]

()

 

[exar333]

please link your source.

no way this is a single gpu card with true 12gb of vram.

if so. that would be significant boost in performance.
in that case. want 4 for 7680x1440.

Quadro K6000 was 12GB. (that was Kepler)

Would not be surprised to see Titan X at 12GB and the ~ Quadro even higher...

 

[.vodka]

I highly, highly doubt we'll see a 1080 out of the GM200. That just makes no sense. The 9xx series launched last fall.

I think we'll see a 985/985Ti or a 990/990Ti. They COULD release a GM200 as a 980Ti, but that would be a weird pairing and a significant leap over the standard 980's GM204.

I suppose we might see a repeat of the 600/700 double generation.

680 -> 770 | 980 -> "1070" would be an excellent way to remove the controversial 970 out of the picture and leave the "1070" to a full GM204, then titan / 780 / 780ti / titan black. Why not? It served them well last time and made a ton of money selling those cut down GK110s up to the full chips, plus giving GK104 some new life.

But then you're right, the 900 brand isn't that old. Yet they haven't used a x90 denomination yet, it could be a first. 980, 990 cut down GM200, 990Ti with a full fat GM200, and leaves them room for a, let's suppose, 980Ti = OCed GM204.

While we're at it: X = 10, or maxwell / tegra X1 :hmm:.

If Nvidia launches the 1080 that fast they may as well rename this the Titan WHY

lol, titan why

 

[nvgpu]

Quadro M6000 will either have 12GB if using current GDDR5 memory chips or 24GB of GDDR5 using the new higher density GDDR5 memory chips from Samsung.

http://techreport.com/news/27676/samsung-starts-making-8gb-gddr5-memory-chips

 

[Headfoot]

I'm really curious to see if they've added even further architectural advancements into Big Maxwell. 750 Ti -> 970/980 added new stuff, I wonder if 980 -> Titan X will too. I would imagine so.

 

[destrekor]

please link your source.

no way this is a single gpu card with true 12gb of vram.

that would be massive if so. might need 4.

First, it was unveiled that it has 12GB of RAM and 8 Billion transistors. If it were 2x980 (2xGM204), it would be over 10 Billion transistors. Heck, even if it were 2xGM206 (GTX 960), it would have 6 Billion transistors, so the number alone is proof it has a different chip.

Also, while it is "a rumor", it is quite clear that the leaked PCB shown below is a GM200 with 12GB of RAM.

[Source]


Now, that specific PCB has 6GB of DDR5, as those are 512MB chips from SKhynix. However, in the 12-chip configuration, which is required to reach a 384bit bus for Nvidia's 32bit Memory Controller configuration (8x32 on GTX 980 = 256bit bus), that means the typical configuration for the GM200 is either going to be 6GB or 12GB. It can easily host 12GB as 12x1GB(8Gb) DRAM chips is readily achievable.

That board pictured above is likely either a test-mule or a 6GB variant destined for a Quadro-brand card.

Really, the 8B transistors fact is the only one that truly matters. That's precisely how large a GM200 should be, and is not a number that makes sense for any dual-GPU configuration.

 

[tviceman]

It will be super mega overpriced i think 1500USD.

I think you're dead on. They might throw a bone and price it at $1200, but anyone who thinks $1200 is a deal is higher than satellite. Final performance will be fun to see in benchmarks but the price will be absurd. The thing is, so long as Nvidia makes a habit out of making a halo-priced GPU, should AMD ever strike gold again (9800 PRO, or great timing a la hd5870), they'll likely follow suit (albeit to a smaller extent). If a $1000 mass-monster video card gets a $400 price drop six months after release, consumers psychologically begin to think it's a good deal. AMD sees this and, while they can hammer on it's price all day long, are envious of not being able to replicate the pricing model.

 

[tviceman]

Interestingly, this is only a ~12.5% transistor count increase over GK110. GM204 had a much larger 49% transistor count increase. It seems the rumors of GM200 not being full FP64 may be true.

 

[metalliax]

Should be a beast for those who need DP and use CUDA.

I'm willing to bet it won't be great for DP. Maxwell was not designed for Double Precision - why else would nVidia release GK210 at the end of 2014?

 

[RussianSensation]

It's shocking to think NV will have the entire Maxwell mobile and desktop generation out before AMD releases a single R9 300 product. As gamers we better hope that R9 300 doesn't flop or the entire Maxwell generation will have high prices until Pascal. Looking forward to Titan X benches and overclocking to see where GM200 consumer card will land. Seeing as to what happened with the $1000 Titan (i.e., 9 months from launch a $400 R9 290 tied it and 1.5 years from launch a $330 GTX970 beat it), I wouldn't buy this card unless I was loaded or needed it for university/work.

NV obviously found a new niche segment where some consumers find it cheaper than buying a full blown workstation/compute Quadro/Tesla card so even at $1500 it would be "cheap" for those intended customers. I wouldn't be surprised at all if the card is $1299-1499.

 

[thilanliyan]

nm

 

[Head1985]

It's shocking to think NV will have the entire Maxwell mobile and desktop generation out before AMD releases a single R9 300 product.

I really want know what happened to AMD.They pretty much start having big troubles after 5870.
5870 destroys GTX285, but after that:
6970 comes after GTX580 and didnt beat it.
7970 beat GTX580, but Nv counter it 3 months later only with mainstream kepler and Big kepler was another league.
Then AMD after 7 months after Big kepler launch hawaii and they didnt even beat kepler.
NV after that launch maxwell and its already 6 month from maxwell launch and we see nothing from AMD and NV even launch Big maxwell faster than AMD Fiji?
And imagine if 380/380x are just renamed 290/290x.
AMD still have only 1.5 year old hawaii vs mainstream maxwell and BIG maxwell.
So whats happened to AMD?No money to develop new GPU or what?

 

[ShintaiDK]

So whats happened to AMD?No money to develop new GPU or what?

Its the painful reality of R&D budgets. Despite that some will claim it doesnt matter and that AMD can still develop GPUs, x86 CPUs, ARM CPUs, etc with a smaller R&D budget than nVidia without any penalty.

The 300 series needs to be a miracle to change anything. But the hybrid cooling leaves an impression that its the last whistle before the game is over like with the 220W FX chips. In a world where performance/watt gets ever more important.

 

[tviceman]

Since, and starting with Tahiti, AMD's major releases have been:

Tahiti
Pitcairn
Cape Verde
Turks
Bonaire
Hawaii
Tonga

(with a smattering of smaller release GPU's like Oland).

Since, and starting with Kepler, Nvidia's major releases have been

GK104
GK107
GK106
GK110
GM107
GM204
GM206

(and a smattering of limited release GK208 and GM207)

Coming soon, GM200.

So if you step back and look at new silicon releases, Nvidia and AMD have released the same number of GPU's up to this point. The difference: AMD has been evolving their current architecture while Nvidia redesigned theirs. The product release cadence has been the same. The difference, though, has been Nvidia being able to stay on top in performance and make much bigger splashes at release.

 

[Railgun]

I'm SWAGging it here in relation to the rumor that it will be gimped in FP64 performance. At the most, I can see $1k...but that's a strech. The only reason the Titan could command that was for it's FP64 performance. There's no other reason. With the 980 currently floating around $550, is it reasonable to say this will be around $850?

Again, it's a SWAG...but I can't see it asking $1k...12G RAM or not.

 

[zlatan]

The difference: AMD has been evolving their current architecture while Nvidia redesigned theirs.

I don't think that building a fully new memory subsystem is just an evolutionary step. It needs an almost complete redesign at the hardware level. And I can tell that building a GPU with 4096-bit data bus is really-really hard. So there is a reason why they didn't announce a new chip.

 

[Gloomy]

You could also say that GCN didn't really need a big redesign. They seem to have hit a home run, so they're iterating on it rather than tearing it down.

Even if you don't agree, it is pretty obvious Kepler is not aging gracefully. So it's no surprise Maxwell had to be a bigger departure.

 

[RussianSensation]

The 300 series needs to be a miracle to change anything. But the hybrid cooling leaves an impression that its the last whistle before the game is over like with the 220W FX chips. In a world where performance/watt gets ever more important.

So many flaws with that statement.

1. We don't know for a fact that Bulldozer is the last CPU architecture from AMD. Implying that FX is the last whistle of AMD's CPU division and disregarding Zen in Q3 2016 is ignorance at its finest. Then you proceed to draw a correlation between AMD's CPU and their GPU divisions as if designing GPUs and CPUs is somehow directly related...:hmm:

2. Implying that R9 300 series' architecture will form a backbone for R9 400 or 500 series is also a theory with no proof -- and this is what you are implying since if you don't believe that, you can't say anything about the efficiency of R9 400 or 500 series since you have no idea what the new architecture underpinning them will be like.

Trying to correlate perf/watt of AMD's R9 300 series made on 28nm towards 14nm/16nm designs is a waste of time unless we know for a fact that AMD's 14nm chips will share similar GCN architecture with R9 300 series. We don't have such data.

Unless you can see the future, I am pretty sure you can't make any inferences whatsoever about 14nm/16nm GPUs from AMD and NV with respect to each other.

"The roadmaps will include AMD’s upcoming products based on the next-generation x86 high-performance Zen core and ARMv8 64bit K12 core. In addition to an all new family of FinFET based GPUs code named Arctic Islands. These will feature the company’s most significant architectural evolution on the GPU front since the introduction of GCN (Graphics Core Next) back in late 2011."
http://wccftech.com/amd-future-gpu-cpu-roadmap/#ixzz3TSC8RWEf

3. 300W TDP + WC tells us nothing that perf/watt of 390X will be horrible as you seem to imply for the last 6 months. In all of your posts you cannot grasp how AMD can possibly improve perf/watt on the same 28nm node so your post is hardly a surprise. However, you have also failed to take into account that a 250W GPU vs. a 300W GPU will get paired with a high-end i5/i7 system, which means a true measure of perf/watt a gamer experiences is Total System perf/watt. Your constant focus on GPU perf/watt and total disregard for total system perf/watt is amusing because no GPU can operate independently of the other system components.

4. R9 300 series doesn't need to win in perf/watt to be a good product. Again, not everyone is a tree hugger. Many people would take a 300W $500 R9 390X with 10% less performance compared to a 250W $700 GM200.

Also, it's interesting how you keep hyping up perf/watt as some savior and a requirement for GPUs yet when AMD owned NV for 3.5 consecutive generations in perf/watt with HD4000-7000 series, they hardly gained market share or made $. Therefore, it can be concluded with 100% certainty that even if AMD were to have class leading perf/watt with R9 300 series, it absolutely does not guarantee market share or financial success either. There are other factors that matter such as marketing, OEM/customer relationships, strong supply chain/manufacturing/logistics, PR, features, timing in getting design wins, etc. that can supersede perf/watt.

Look at the NV shield console. It will hands down have the best perf/watt among Wii U, PS4/XB1 but that console is going to bomb in sales compared to any of them. Perf/watt is not some magical metric that guarantees success. NV is successful not only because of perf/watt but many other factors.

 

[Deders]

12GB v ram hmm it will be 100% GM200.
3072SP
192TMU
96Rops
384bit
12GB Vram
performance 50% over GTX980.

It will be super mega overpriced i think 1500USD.

Yep, the 8 billion transistor count compared to 5.2 billion points to this. I just hope they don't cut down the consumer version too much so it's uneven like the 970.

 

[Subyman]

I'm SWAGging it here in relation to the rumor that it will be gimped in FP64 performance. At the most, I can see $1k...but that's a strech. The only reason the Titan could command that was for it's FP64 performance. There's no other reason. With the 980 currently floating around $550, is it reasonable to say this will be around $850?

Again, it's a SWAG...but I can't see it asking $1k...12G RAM or not.

I don't see this thing being under $1k, unless its $999.99. They have absolutely no competition in that performance range.

 

[tviceman]

I don't think that building a fully new memory subsystem is just an evolutionary step. It needs an almost complete redesign at the hardware level. And I can tell that building a GPU with 4096-bit data bus is really-really hard. So there is a reason why they didn't announce a new chip.

Of course anything new is hard. What about 8 billion transistors with GM200? You think that is easy?! AMD is out of excuses and has no room for further mistakes. They're less than 25% market share in discrete graphics and are selling their Hawaii-based cards at nearly half the price of the smaller, less expensive GM204 gtx980. They need Fiji to be great and sell like it. Fiji needs to be very competitive with GM200 on all fronts.

 

[Cloudfire777]

OMG, this may be my next card!!!!

I hope it got 2816 cores. Please let them unleash a full GM204 straight off the bat

 

[Silverforce11]

Interestingly, this is only a ~12.5% transistor count increase over GK110. GM204 had a much larger 49% transistor count increase. It seems the rumors of GM200 not being full FP64 may be true.

GM204 was a lot bigger than GK104 though. They don't have much room to grow from GK110, ~600mm2 is the limit at TSMC.

I'm expecting ~35% faster than 980 at ~780ti/Titan power levels (~220W), but they do have plenty of TDP room to clock it higher at stock for extra performance.