New Titan X announced, 12GB GDDR5X, $1200

Status
Not open for further replies.
The thing looks like a beast which is awesome. But keeping the same name is going to cause a lot of resale issues on auction sites. New better product should have an addition to the name. For crying out loud at least call it Titan X2 or something.
 
Dayaks said:
Just realize there will be HBM2 cards next year and it wouldn't surprise me if there's a 600mm^2 die. Get those two things and I believe we're out of tech jumps for a long, long time.
Click to expand...

There likely will be a 600mm GPU on this node, maybe next year with a Pascal refresh, or maybe it'll just be Volta on the same process. Either way, I don't see why HBM matters, it's only useful when you need the bandwidth and/or power savings, both of which aren't needed right now with g5x
 
Ieldra said:
There likely will be a 600mm GPU on this node, maybe next year with a Pascal refresh, or maybe it'll just be Volta on the same process. Either way, I don't see why HBM matters, it's only useful when you need the bandwidth and/or power savings, both of which aren't needed right now with g5x
Click to expand...
Yep, this feels like the 400 series days where a 11 series will be a refresh with a fully enabled big chip. I just wish it was 400 series pricing....
 
To put it more precise, GDDR5X killed HBM2 for gaming. HBM2 for compute however is still needed, this is also why we see it on GP100. If AMD didn't have HPC dreams with Vega, I doubt they use HBM2 at all. HBM was an outright disaster for Fiji for example. GDDR5X will also keep scaling, just with 14Ghz GDDR5X in the future, bandwidth is up another 40% for any refresh cards.
 
Stev3FrencH said:
Yep, this feels like the 400 series days where a 11 series will be a refresh with a fully enabled big chip. I just wish it was 400 series pricing....
Click to expand...

Meh, the pricing isn't too bad honestly, price gouging aside. NV sets the MSRP, and it really isn't too shabby frankly; cost increase per transistor for the first time ever on this new node + effective 50% transistor increase comparing gm200 to Gp102 and we're talking about a die that's *at least* ~60% more expensive to produce. It's also about 6 months ahead of its competition. $1200 is a fine price for this kind of halo product
 
Shintai said:
TO put it more precise, GDDR5X killed HBM2 for gaming. HBM2 for compute however is still needed, this is also why we see it on GP100.
Click to expand...
Because of DP and the neural networking applications mainly, also relatively small power savings stack up fast when you're running racks upon racks of DGX-1s; and the cost of HBM is amortized by the 120$k price tag :p
 
Ieldra said:
Because of DP and the neural networking applications mainly, also relatively small power savings stack up fast when you're running racks upon racks of DGX-1s; and the cost of HBM is amortized by the 120$k price tag :p
Click to expand...

With the DGX-1 I am sure size was also a an issue, tho tiny. But yep, HBM2 for compute bandwidth and ECC. The power savings are really minimal if you look beyond PR slides and got no real impact. Its one that got into the hype area.

GP100 and GP102 shows how the cost structure is. 450 vs 600mm2 die and GDDR5X vs HBM2. This was so great that it was cheaper to spend 100M$ or more to make a new chip.
 
Shintai said:
With the DGX-1 I am sure size was also a an issue, tho tiny. But yep, HBM2 for compute bandwidth and ECC. The power savings are really minimal if you look beyond PR slides and got no real impact. Its one that got into the hype area.

GP100 and GP102 shows how the cost structure is. 450 vs 600mm2 die and GDDR5X vs HBM2. This was so great that it was cheaper to spend 100M$ or more to make a new chip.
Click to expand...

Well the real cost savings come from gutting fp64 and halving the register file size. Doubling the ALUs/SM (all things equal) will decrease chip size.

Anyway, it was inevitable that there would be Gp102 because of the Quadro line; they're not gonna sell a GPU with 1:2 FP64 ratio to anyone who doesn't pay $5.5k :p
 
Dayaks said:
Just realize there will be HBM2 cards next year and it wouldn't surprise me if there's a 600mm^2 die. Get those two things and I believe we're out of tech jumps for a long, long time.
Click to expand...

Yeah, I spoke before I realised the 'no HBM2' part.

That changes everything. Not worth it at $1,200. I'd pay $1,000 for this, and up to $1,500 for a HBM2 variant, but I don't think I will as it is.

Such a shame! :rolleyes:
 
spine said:
Yeah, I spoke before I realised the no HBM2.

That changes everything. Not worth it at $1,200. I'd pay $1,000 for this, and up to $1,500 for a HBM2 variant, but I don't think I will as it is.

Such a shame! :rolleyes:
Click to expand...

Why would you pay more for HBM2 if it doesn't add anything useful? Because its "new"?

The question today is if HBM will ever take off. Or if GDDR5X and future incarnations nailed HBM for good and HMC takes over the top.
 
Shintai said:
Why would you pay more for HBM2 if it doesn't add anything useful? Because its "new"?

The question today is if HBM will ever take off. Or if GDDR5X and future incarnations nailed HBM for good and HMC takes over the top.
Click to expand...

Yeah I love this lol, 1200 is too much for this! I'd only pay 1000,but I'd gladly shell out another 300 for a feature I'm likely never going to see the benefit of. Makes perfect sense :p
 
I don't know if I believe it's not bandwidth constrained at all right now. Do you have data to say it's not?

It takes a hell of a lot of bandwidth in some of these games. People OC their VRAM for a reason? (Could be just because the slider is there...)
 
Dayaks said:
I don't know if I believe it's not bandwidth constrained at all right now. Do you have data to say it's not?

It takes a hell of a lot of bandwidth in some of these games. People OC their VRAM for a reason? (Could be just because the slider is there...)
Click to expand...

Course there's data! Go look at GP104, is it memory limited? By how much? Gp102 is going to be less memory limited, because it has 7% fewer SMs to feed per unit bandwidth.


Best way to test memory limitation is; benchmark at stock, the bench with only core OC, then bench with core + memory OC
 
Ieldra said:
Course there's data! Go look at GP104, is it memory limited? By how much? Gp102 is going to be less memory limited, because it has 7% fewer SMs to feed per unit bandwidth.


Best way to test memory limitation is; benchmark at stock, the bench with only core OC, then bench with core + memory OC
Click to expand...

There's data on an overclocked, power limit unlocked, GP104?
 
Dayaks said:
I don't know if I believe it's not bandwidth constrained at all right now. Do you have data to say it's not?

It takes a hell of a lot of bandwidth in some of these games. People OC their VRAM for a reason? (Could be just because the slider is there...)
Click to expand...

I think Nvidia could easily by now fit it with 12Ghz GDDR5X if it was that bad. I am sure Micron can deliver that today. The Titan got more bandwidth in terms of everything else than the 1080 for example due to core clocks. So the Titan will be less bottlenecked than the 1080. So the question is, how bottlenecked is the 1080 on memory?
 
Shintai said:
I think Nvidia could easily by now fit it with 12Ghz GDDR5X if it was that bad. I am sure Micron can deliver that today. The Titan got more bandwidth in terms of everything else than the 1080 for example due to core clocks. So the Titan will be less bottlenecked than the 1080. So the question is, how bottlenecked is the 1080 on memory?
Click to expand...

Yeah my GP102 predictions were 30SMs, 6GPCs, 240 TMUs, 96ROP and 12GT/s memory on 384b bus :( which is actually spot on, except for the 10GT/s memory
 
Dayaks said:
There's data on an overclocked, power limit unlocked, GP104?
Click to expand...
All you need is to look at how memory OC affects performance at a given (high clock).

So say you have a 2ghz 1080 performing at 100%.

A 10% memory OC gives you 3% extra performance, with Gp102 the performance gain will be lower by about 5-10%
 
Shintai said:
Why would you pay more for HBM2 if it doesn't add anything useful? Because its "new"?
Click to expand...

Ieldra said:
Yeah I love this lol, 1200 is too much for this! I'd only pay 1000,but I'd gladly shell out another 300 for a feature I'm likely never going to see the benefit of. Makes perfect sense :p
Click to expand...

You're both right, I'm being dumb. But I just feel like spending a shit ton of cash on myself right now on something I don't need.

That's the market for these cards, and this time around, I might actually be in it! :p
 
Also worth mentioning AMD have already bitten themselves in the ass by hyping HBM time and time again.

HBM will hinder their ability to respond to GP104 and GP102 in good time.

HBM's power savings will be smaller than on 28nm.

HBM's bandwidth alone won't give them an edge over the competing Pascal parts.

If they release Vega 10 with G5X all the AMD fans they fooled into thinking HBM is the most important thing since sliced cheese will be disappointed and will rage.

Now Roy Taylor is making sneaky comments about waiting for Christmas, in response to the Titan announcement. All aboard the hype train, choo choo!
 
Ieldra said:
All you need is to look at how memory OC affects performance at a given (high clock).

So say you have a 2ghz 1080 performing at 100%.

A 10% memory OC gives you 3% extra performance, with Gp102 the performance gain will be lower by about 5-10%
Click to expand...

But it has to be power limit unlocked and in actual games to be valid to me. 12GB of VRAM is a lot to feed over that bus. So high fps games with high VRAM usage would be worst case. Most people just look at firestrike which uses very little VRAM.
 
Ieldra said:
Also worth mentioning AMD have already bitten themselves in the ass by hyping HBM time and time again.

HBM will hinder their ability to respond to GP104 and GP102 in good time.

HBM's power savings will be smaller than on 28nm.

HBM's bandwidth alone won't give them an edge over the competing Pascal parts.

If they release Vega 10 with G5X all the AMD fans they fooled into thinking HBM is the most important thing since sliced cheese will be disappointed and will rage.

Now Roy Taylor is making sneaky comments about waiting for Christmas, in response to the Titan announcement. All aboard the hype train, choo choo!
Click to expand...

4GB of HBM was stupid and most us called it out before the card even launched.

Now if you're talking about 16GB and high Hz, that's a different animal.
 
Dayaks said:
But it has to be power limit unlocked and in actual games to be valid to me. 12GB of VRAM is a lot to feed over that bus. So high fps games with high VRAM usage would be worst case. Most people just look at firestrike which uses very little VRAM.
Click to expand...
VRAM contains much more than just the frame buffer, and I don't see how the amount of memory is relevant as far as bandwidth is concerned.

High fps, low fps, doesn't matter, the question is whether the bandwidth will choke the SMs or not, and the answer is; around 8% less than on GP104.

You don't need to be using a lot of VRAM to be maxing out memory bandwidth and choking the GPU.

I dont know if it was here or on Guru3D, but some guy went on a rant about how Titan X is 'gimped' (the maxwell one) by its memory bandwidth. So obviously, I asked him what he was on about.
He said the card is designed not to even hold 60fps.

What? Weird, right?

Anyway, so I asked him how he reached that conclusion:

" it's obvious, 12GB VRAM, 60 frames per second, 720GB/s required "

Lol
 
Pretty soon you'll have to go to a dealer to buy one so you can secure 36 month financing.

$1200 for a video card that will be obsolete in under 12 months. No thanks.
 
Ieldra said:
VRAM contains much more than just the frame buffer, and I don't see how the amount of memory is relevant as far as bandwidth is concerned.

High fps, low fps, doesn't matter, the question is whether the bandwidth will choke the SMs or not, and the answer is; around 8% less than on GP104.

You don't need to be using a lot of VRAM to be maxing out memory bandwidth and choking the GPU.

I dont know if it was here or on Guru3D, but some guy went on a rant about how Titan X is 'gimped' (the maxwell one) by its memory bandwidth. So obviously, I asked him what he was on about.
He said the card is designed not to even hold 60fps.

What? Weird, right?

Anyway, so I asked him how he reached that conclusion:

" it's obvious, 12GB VRAM, 60 frames per second, 720GB/s required "

Lol
Click to expand...

ROFL!!!
 
If I get to power limit unlocking my card today I might do some benches for funsies. I figure the PL needs to be gone or you're not changing just once variable...
 
Ieldra said:
Tell us, prophet Riccochet, why will it be obsolete in less than a year?
Click to expand...

Not obsolete in the sense that it won't be useful, but obsolete in terms of being a performance leader. In 12 months a new card will come, cost half as much and offer slightly better performance. Which makes paying $1200 for a Titan kind of pointless. Unless you have money to burn to be on the bleeding edge. I don't. I bass fish. I'm lucky I have money for toilet paper and food. ;)
 
Riccochet said:
Not obsolete in the sense that it won't be useful, but obsolete in terms of being a performance leader. In 12 months a new card will come, cost half as much and offer slightly better performance. Which makes paying $1200 for a Titan kind of pointless. Unless you have money to burn to be on the bleeding edge. I don't. I bass fish. I'm lucky I have money for toilet paper and food. ;)
Click to expand...
Nah. Not happening man. Best case scenario the next lineup of cards will give you performance of this Titan X for 1080 prices.

It's definitely expensive, and its not meant to be cost efficient, certainly isn't for nvidia either :p

Dayaks said:
If I get to power limit unlocking my card today I might do some benches for funsies. I figure the PL needs to be gone or you're not changing just once variable...
Click to expand...

You could just settle on a clock that doesn't make you hit PL and compare performance with and without memory OC
 
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.
 
4k games @60hz maxed will probably be the draw here. Otherwise the price/performance value is kinda crappy. The good tidbit is that there is likely going to be a 1080Ti to follow shortly after this card. Maybe at a $899.99 price point and that will make alot of hardcore enthusiasts happy. I'm surprised Nvidia is running straight into the deep this generation. Usually they test the watters and go about knee deep and wait and see what AMD will bring and if it's worth releasing the big guns. I'm convinced that they will basically create a more refined Pascal 2.0 version of the 1080 and 1080 Ti with higher clocks, lower temps and release those with HBM 2 stacked ram in early-mid 2017. Maybe the next Gen Volta in early 2018.

Hmm. All I know is that 1080 P is what I use my card for since I stream and there is no need for anything more, so I wont need to upgrade for a very very long time.
 
It's seems clear Vega will release this year, and I can't see that nvidia wouldn't respond with something, anything, at that time, even if Vega fails to match a regular 1080.

So I expect we'll see a 1080ti later this year now. I was previously expecting this Titan to land with Vega, but I guess Nvidia saw an opportunity over the summer to grab some extra cash!

I mean, you gotta hand it to nvidia, they are milking this market exceptionally well!
 
Ieldra said:
Nah. Not happening man. Best case scenario the next lineup of cards will give you performance of this Titan X for 1080 prices.

It's definitely expensive, and its not meant to be cost efficient, certainly isn't for nvidia either :p
Click to expand...

Really now? Lets see. Titan X was released March of 2015. 980 ti in June 2015. A year later we have 1080 that's faster than Titan X. Please explain to me how "Nah. Not happening man". And remember, 980 ti when overclocked was as fast as Titan X, for the most part. And it came out ~3 months after the Titan X.
 
The clock speed on the box is basically irrelevant as the GPU will probably click higher than that on its own. You'd have to be mad to buy this and keep the stock cooler frankly.

Riccochet

The 980ti was much faster than a stock Titan X when overclocked, it's also had half the VRAM though. With this Titan there's little space for another cut because it's going to be too close to the 1080. They can't sell a Ti with the same SM count because vram is only 12gb and they cant halve it.

An overclocked 980ti still holds its own against a 1080. Don't see the issue really, if it bothers you that you'll always get better performance for the same price one year down the line you should really find a new hobby
 
Lord_Exodia said:
Maybe at a $899.99 price point and that will make alot of hardcore enthusiasts happy.
Click to expand...

I've been an enthusiast for a very long time, but $899...is $899.
Both the 780/980Ti remain the most I've spent on a card ($650, $600 respectively). nVidia gear tends to be better values *only* after price cuts, and encouraged by competitive AMD hardware.

Overall - I just don't see myself spending that much money. It's really getting ridiculous.
 
Mav451 said:
I've been an enthusiast for a very long time, but $899...is $899.
Both the 780/980Ti remain the most I've spent on a card ($650, $600 respectively). nVidia gear tends to be better values *only* after price cuts, and encouraged by competitive AMD hardware.

Overall - I just don't see myself spending that much money. It's really getting ridiculous.
Click to expand...


This card is going to sell for HPC, having it cheaper would just make it even harder to find in stock. Titan cards were always lacked cost efficiency for gaming.

Titan X was a very cost efficient card for neural networks. Terrible cost efficiency for gaming

New Titan X will be the same
 
Ieldra said:
The clock speed on the box is basically irrelevant as the GPU will probably click higher than that on its own. You'd have to be mad to buy this and keep the stock cooler frankly.

Riccochet

The 980ti was much faster than a stock Titan X when overclocked, it's also had half the VRAM though. With this Titan there's little space for another cut because it's going to be too close to the 1080. They can't sell a Ti with the same SM count because vram is only 12gb and they cant halve it.

An overclocked 980ti still holds its own against a 1080. Don't see the issue really, if it bothers you that you'll always get better performance for the same price one year down the line you should really find a new hobby
Click to expand...

Mav451 said:
I've been an enthusiast for a very long time, but $899...is $899.
Both the 780/980Ti remain the most I've spent on a card ($650, $600 respectively). nVidia gear tends to be better values *only* after price cuts, and encouraged by competitive AMD hardware.

Overall - I just don't see myself spending that much money. It's really getting ridiculous.
Click to expand...

That's the point I'm trying to make. This new Titan at $1200 is just stupid.

And, like I said, I bass fish. Makes building custom, high end PC's look like childs play in terms of cost.

Everyone is entitled to their opinion and where they choose to spend their money for the best "bang4buck".
 
I just figured out the naming scheme for this GPU. It is Titan X, not Titan X, like the previous part. They did that to differentiate the two Titan X parts.


The original Titan X was TITAN + the letter X, this one is TITAN + the Roman numeral X (Power of 10). Makes total sense.
 
Status
Not open for further replies.
Back
Top