New Titan X announced, 12GB GDDR5X, $1200

Status
Not open for further replies.
viscountalpha said:
HBM2 might not be ready for the mainstream
Click to expand...

And it will never be. Vega+GP100+GV100 looks to be all GPU cards HBM2 will ever be on. HBM is flawed in its basis in terms of cost structure as well. It will never be cheap. Its also interesting that AMD says next gen memory and not something like HBM3. Most likely GDDR6.
 
Guys. I think I figured this shit out.

This *is* the 1080ti.

This will be rebranded as 1080ti and AIB will do custom boards and cooling

Full 30SM 3840 ALU GP102 with 24GB g5x (possibly 12GT/s) will take its place as Titan

Either this or they'll play it like Kepler
 
Last edited:
ecmaster76 said:
Going by the Anandtech blurb even though this has many more SM it is significantly lower clocked. Total peak compute is ~22% higher for 2x the price. The extra VRAM and bandwidth might add a few more percent but unless this thing has some extra secret sauce vs the 1080 then it will have pretty much the worst price/performance ratio of the current lineup by far.

I know these are supposed to be ultra enthusiast halo cards but this is looking more like a cheap Quadro than anything.
Click to expand...

This was my feeling as well. I'm curious to know if it works well as a Quadro replacement or not.

I still expect them to make a 1080 ti in 6 to 12 months.
 
viscountalpha said:
This was my feeling as well. I'm curious to know if it works well as a Quadro replacement or not.

I still expect them to make a 1080 ti in 6 to 12 months.
Click to expand...
Just to remind you. Stock boost on original Titan X and 980Ti was 1075 mhz. We all know how that worked out.

This card will probably clock to 2ghz easily, just needs adequate cooling
 
Damn, no more magnesium alloy body? The GeForce page says the body is cheaper die-cast aluminum. That's a shame.

NVIDIA TITAN X Graphics Card with Pascal | GeForce
ecmaster76 said:
Going by the Anandtech blurb even though this has many more SM it is significantly lower clocked. Total peak compute is ~22% higher for 2x the price. The extra VRAM and bandwidth might add a few more percent but unless this thing has some extra secret sauce vs the 1080 then it will have pretty much the worst price/performance ratio of the current lineup by far.

I know these are supposed to be ultra enthusiast halo cards but this is looking more like a cheap Quadro than anything.
Click to expand...
That is what the numbers on paper say. Also, the Titan series has typically always been Quadro rejects. This Titan just feels more like a bigger GP104, though.
Shintai said:
And it will never be. Vega+GP100+GV100 looks to be all GPU cards HBM2 will ever be on. HBM is flawed in its basis in terms of cost structure as well. It will never be cheap. Its also interesting that AMD says next gen memory and not something like HBM3. Most likely GDDR6.
Click to expand...
I love this. There is no "next gen" memory. The slide in AMD's presentation said "Nexgen" memory. Nexgen was a semiconductor company that AMD acquired a long time ago, but went defunct in 1996. There is a company out there called NexGen that puts flash memory in the PCI-E slot, and I think this is what AMD is referring to. This is the product page for their current architecture, but this is certainly not practical for the consumer. They're most likely working on something specific for AMD.

NexGen :: Architecture
Ieldra said:
Guys. I think I figured this shit out.

This *is* the 1080ti.

This will be regarded 1080ti and AIB will do custom boards and cooling

Full 30SM 3840 ALU GP102 with 24GB g5x (possibly 12GT/s) will take its place as Titan

Either this or they'll play it like Kepler
Click to expand...
No custom cooling or boards have ever been allowed with a Titan chip. The Hybrid from EVGA with the last Titan X was the exception.

Pascal is probably going to have a short shelf life, and Volta is probably going to be launched in the fall (northern hemisphere...) next year.
 
Armenius said:
That is what the numbers on paper say. Also, the Titan series has typically always been Quadro rejects. This Titan just feels more like a bigger GP104, though.

I love this. There is no "next gen" memory. The slide in AMD's presentation said "Nexgen" memory. Nexgen was a semiconductor company that AMD acquired a long time ago, but went defunct in 1996. There is a company out there called NexGen that puts flash memory in the PCI-E slot, and I think this is what AMD is referring to. This is the product page for their current architecture, but this is certainly not practical for the consumer. They're most likely working on something specific for AMD.

NexGen :: Architecture

No custom cooling or boards have ever been allowed with a Titan chip. The Hybrid from EVGA with the last Titan X was the exception.

Pascal is probably going to have a short shelf life, and Volta is probably going to be launched in the fall (northern hemisphere...) next year.
Click to expand...

I'm saying this 3584 Titan X will become the 1080ti with custom boards and coolers from AIB.

Another Titan will take its place with 24gb and 3840 alu
 
Ieldra said:
I'm saying this 3584 Titan X will become the 1080ti with custom boards and coolers from AIB.

Another Titan will take its place with 24gb and 3840 alu
Click to expand...

Starting at the low low MSRP of $1099, or you know, $1149
 
Dayaks said:
I was hoping for a full 600mm^2 chip on this node and then tiny increments gains afterwards (think Intel). They could have done a gaming GPU at 600mm^2 if they wanted...
Click to expand...

I prefer performance over die-space if we should select metrics...but that would be a retarded exercise comparing different metrics.

Yours are useless btw since NVIDIA copied the same 3:2:1 SKU as they did with Maxwell.

A larger die would not give more performance, but it would cost you more...and if I need to explain why...you seriously need to stop posting.
 
Hope AIB cards will come out shortly after launch but it might not happen because for some reason NVIDIA forbids their AIB's to create custom coolers for their Titan cards, but still curious to see how this card will perform at 4K.
 
Hmm... Do you guys think there'll even be a GTX 1080ti?

I initially thought there would be one and it would probably cost like $899? (totally expected Nvidia to be selling the next Titan at $1200-1300.)

I just don't know how they're going to gimp the GP102 to make a GTX1080ti. I mean... they can't just make a GPU102 8GB VRAM right? they'd have to gimp it down to 6GB no? If they did that it would be weird that the GTX 1080 has 8GB, then again GP102 is running on 384 bits bus so there's that...

(sorry my post is highly speculative... Just want to see what you all think.)
 
Flexion said:
Hmm... Do you guys think there'll even be a GTX 1080ti?

I initially thought there would be one and it would probably cost like $899? (totally expected Nvidia to be selling the next Titan at $1200-1300.)

I just don't know how they're going to gimp the GP102 to make a GTX1080ti. I mean... they can't just make a GPU102 8GB VRAM right? they'd have to gimp it down to 6GB no? If they did that it would be weird that the GTX 1080 has 8GB, then again GP102 is running on 384 bits bus so there's that...

(sorry my post is highly speculative... Just want to see what you all think.)
Click to expand...

A 1080TI will be 12GB, not any question about that. As yield improves and Titan X have sold its initial batch I'm sure we will see it. Perhaps closer to the holidays. November or so.
 
Factum said:
I prefer performance over die-space if we should select metrics...but that would be a retarded exercise comparing different metrics.

Yours are useless btw since NVIDIA copied the same 3:2:1 SKU as they did with Maxwell.

A larger die would not give more performance, but it would cost you more...and if I need to explain why...you seriously need to stop posting.
Click to expand...

Go on. They used a 600+ mm^2 die for P100, why not for 102?
 
Dayaks said:
Go on. They used a 600+ mm^2 die for P100, why not for 102?
Click to expand...
May have to do with them not caring one minute about yields on GP100, since it's the part they can price almost as high as they want to (bargain bin DGX-1, anyone?).
 
I agree with the other poster who thinks this will become the Ti. Down the road, the Titan will likely be revved in about 9 months and have double the memory and a slightly larger die + clock speed with custom cooling solution. Then they can call it the Titan X2 or XL and sell it for 1400.
 
Dayaks said:
Go on. They used a 600+ mm^2 die for P100, why not for 102?
Click to expand...
No FP64, most likely (like Mitan X). It's also not going to be as big as GM200 due to the node shrink. Smaller die also means potentially higher yields.
 
There is so much misinformation about HBM. It's an insane technology.. and if you know anything about RAM, you know that it is often a bottleneck for the corresponding core.

NV didn't choose GDDR5X over HBM because it wasn't "that" good - it was because it has a vastly more expensive manufacturing process. It is extremely non-traditional, and you can kill an expensive die by performing the interposer process.

Please, stop touting "HBM wasn't so great", "HBM is not good enough for gaming", "Only 4GB of HBM with Fury X was awful" when the fact of the matter is it was only just expensive. The performance is there.

If it sucked so much, they wouldn't be using HBM2 on GP100 for real machine learning/deep learning applications.
 
Dayaks said:
Go on. They used a 600+ mm^2 die for P100, why not for 102?
Click to expand...

To me it has something to do with the loads and heat dissipation. P100 has lots of FP units and its TDP is 300 W for its die size and
maximum load. TX has graphics computing units, has higher clock and TDP of 250 W achieved already at smaller die size. Since one can not
have TDP more than 250-300 W, the TX die size is (almost) maximum. I expect there will be another Monster Titan X with slightly more graphics
units in bigger die, TDP 300 W, HBM2 memory and the price people die when see it :android:.
 
junkman_ said:
There is so much misinformation about HBM. It's an insane technology.. and if you know anything about RAM, you know that it is often a bottleneck for the corresponding core.

NV didn't choose GDDR5X over HBM because it wasn't "that" good - it was because it has a vastly more expensive manufacturing process. It is extremely non-traditional, and you can kill an expensive die by performing the interposer process.

Please, stop touting "HBM wasn't so great", "HBM is not good enough for gaming", "Only 4GB of HBM with Fury X was awful" when the fact of the matter is it was only just expensive. The performance is there.

If it sucked so much, they wouldn't be using HBM2 on GP100 for real machine learning/deep learning applications.
Click to expand...

There's HBM on GP100 because of the DP, which is a memory hog + power savings overall in DGX-1 and lower access times which are useful considering hybrid cube mesh configuration of the GP100 parts using nvlink. Makes perfect sense for this product, makes no sense at all for a gaming card


As far as games are concerned, we're not the point where games require GPUs to be powerful enough to saturate the 500GB/s achievable with g5x on 384b
 
Massamune said:
I agree with the other poster who thinks this will become the Ti. Down the road, the Titan will likely be revved in about 9 months and have double the memory and a slightly larger die + clock speed with custom cooling solution. Then they can call it the Titan X2 or XL and sell it for 1400.
Click to expand...

Lol that is just wishful thinking. Titan x2 10% faster than this Titan x? This is going to become the ti? Just doesn't make any sense to be honest. And not happening. 24gb of memory? Insane not needed for a gaming product, not now!
 
NKD said:
Lol that is just wishful thinking. Titan x2 10% faster than this Titan x? This is going to become the ti? Just doesn't make any sense to be honest. And not happening. 24gb of memory? Insane not needed for a gaming product, not now!
Click to expand...
Dude... Titan isn't a gaming product.

24GB makes perfect sense for neural networking applications, hence the 24GB Quadro m6000.

What currently sells as Pascal Titan X could become the Ti, with the full GP102 die with 24gb memory taking its place as the Titan X
 
Ieldra said:
There's HBM on GP100 because of the DP, which is a memory hog + power savings overall in DGX-1 and lower access times which are useful considering hybrid cube mesh configuration of the GP100 parts using nvlink. Makes perfect sense for this product, makes no sense at all for a gaming card


As far as games are concerned, we're not the point where games require GPUs to be powerful enough to saturate the 500GB/s achievable with g5x on 384b
Click to expand...


You are correct, but the manufacturing process cost is more important. For HBM1, if you had 1000 and ~90% (900 remain) yield on die, and then combined with HBM for the extra cost and an additional ~30-40% (~600 remain) yield hit for the interposer - the amount of die you lose is substantial. In further addition, there are no reclaimable die after you break one with HBM. Now, imagine that with a GPU there are even less die per wafer (~50-200 maybe), and with FinFET yields are still not mature. You must jack up the price to substantiate that yield loss.

But, you are correct about the consumption savings and performance.
 
Last edited:
nVidia touted the Titan X as a gaming card, IIRC. It does seem that Titan XL or 2.0 or Pitan X is being aimed at the 'prosumer' and not hardcore gamer. I'm not in the market here (anything above the 1070 is well above my budget) but this is a very interesting card if I were looking for a part-time gaming and part-time CAD/CAM/Science card.
 
Ieldra said:
Dude... Titan isn't a gaming product.

24GB makes perfect sense for neural networking applications, hence the 24GB Quadro m6000.

What currently sells as Pascal Titan X could become the Ti, with the full GP102 die with 24gb memory taking its place as the Titan X
Click to expand...
Titan is a GeForce product, and GeForce is NVIDIA's gaming lineup. Even Titan Z is marketed as a gaming video card. Pitan X is a weird one, though, because there is no GeForce labeling on it yet it is listed on GeForce.com under the GeForce product lineup.
 
Armenius said:
Titan is a GeForce product, and GeForce is NVIDIA's gaming lineup. Even Titan Z is marketed as a gaming video card. Pitan X is a weird one, though, because there is no GeForce labeling on it yet it is listed on GeForce.com under the GeForce product lineup.
Click to expand...

Yeah, I wasn't really talking about what it's marketed as, or what it's labelled (GeForce/Quadro/Tesla). Mitan X was clearly GeForce, but it was NOT cost efficient as a gaming card, sure it was great, but it wasn't exactly flying off shelves because of those few gamers who would purchase such a card.
 
Somewhere inside NVidia there must be employees laughing their ass off at this pricing.

"We keep jacking it up, and they just keep paying".
 
Ieldra said:
Why does this have half the memory it's supposed to :p are g5x chips not available in 16gbit?
Click to expand...

Yeah, new tech is new? A little demanding, aren't we :D

Nvidia is selling this Titan X with uncut shaders. Nvidia is getting BETTER yields on massive P100 on a cutting-edge process than HBM2 stacks :D

But this is interesting. I didn't think they could fit the logic for dual memory controllers on P100, so I always assumed the Titan part would be HBM2.
 
Ieldra said:
Dude... Titan isn't a gaming product.

24GB makes perfect sense for neural networking applications, hence the 24GB Quadro m6000.

What currently sells as Pascal Titan X could become the Ti, with the full GP102 die with 24gb memory taking its place as the Titan X
Click to expand...

That was exactly my point. When was the last time Nvidia did that though. Made Titan x and then made that in to a ti and then released another Titan x with another 10% more shaders? And double the memory? Unless the believe people are stupid. I just don't see them doing it. Just sounds like a mess.Lol
 
lolfail9001 said:
This Titan is GDDR5X so HBM2 yields are irrelevant.
Click to expand...

No, you're right. After seeing the cut 12GB Teslas, I just assumed they did the same here, my mind automatically filtered the "GDDR5X" part of the title :D

I'm amazed that they can handle both memory types on the same die, and still have room for all that compute. Doesn't memory controller interface take up a sizable portion of the die? While GDDR5 and GDDR5x are pretty close in signaling and interface, you'd think the lower-speed wider bus of HBM wouldn't work the same?
 
NKD said:
That was exactly my point. When was the last time Nvidia did that though. Made Titan x and then made that in to a ti and then released another Titan x with another 10% more shaders? And double the memory? Unless the believe people are stupid. I just don't see them doing it. Just sounds like a mess.Lol
Click to expand...

How is it any different from releasing a Titan card, then following it up 3 months later with the same card, half the VRAM and ~2 SMs less ?

You buy Pitan X today. 12GB, 3584 SP $1200

Dec 2016: 1080ti: 12GB 3584 SP $850
Dec 2016: Pitan X : 24GB 3584 SP $1200

This is exactly the same as Mitan X -> 980Ti from the viewpoint of a Mitan X owner
 
defaultluser said:
No, you're right. After seeing the cut 12GB Teslas, I just assumed they did the same here. I'm amazed that they can handle both memory types on the same die, and still have room for all that compute. Doesn't memory controller interface take up a sizable portion of the die?
Click to expand...
This is a totally different die, using different ipblocks entirely.

GP100 uses CC6.0 ; 64 ALU/SM, HBM IMC

GP102: CC6.1 ; 128 ALU/SM, GDDR IMC
 
Ieldra said:
This is a totally different die, using different ipblocks entirely.

GP100 uses CC6.0 ; 64 ALU/SM, HBM IMC

GP102: CC6.1 ; 128 ALU/SM, GDDR IMC
Click to expand...

OKAY, that makes sense. I was wondering where the hell GP102 ran off to, and now it's pretty clear where it went.
 
defaultluser said:
OKAY, that makes sense. I was wondering where the hell GP102 ran off to, and now it's pretty clear where it went.
Click to expand...
GP100 - Tesla, some Quadros
GP102/104/106 : Geforce, Quadro, some Tesla, and drive px2
 
Dayaks said:
Go on. They used a 600+ mm^2 die for P100, why not for 102?
Click to expand...

You just want to "tack on" SM to an existing SKU...excuse me while I do a face to desk.

Hint:

Look into how NVIDIA made 3:2:1 SKU's with first Maxwell and now Pascal.
 
Ieldra said:
GP100 - Tesla, some Quadros
GP102/104/106 : Geforce, Quadro, some Tesla, and drive px2
Click to expand...

So does GP102 have thew same Double Precision beef of the P100? Nobody seems to have any hard data on that, but I'm guessing not.

If not, then it's clear what the target market here is: a MUCH CHEAPER TO BUILD P100 suitable for most compute tasks :D
 
Status
Not open for further replies.
Back
Top