Titan X Launch

[Vesku]

What are the chances there will still be a 980ti falling somewhere between this and the 980?

Fairly high chance, having 12GB on this Titan X gives some leeway for releasing a more reasonably priced 6GB version later on.

Compute is gone?
http://www.anandtech.com/show/9059/the-nvidia-geforce-gtx-titan-x-review/15
I beg to differ. It's actually much improved over original Titan save Dual Precision.

Thermally limited? Explain. All GPUs these days are thermally limited.

Titan and Titan Black were marketed as a prosumer card offering DP performance previously reserved for workstation cards. The Titan X has dropped that feature, not something that should just be glossed over, it's a Titan not a 980Ti/985/990.

As for the thermals, from Anandtech's review:

Speaking of clockspeeds, taking a look at our average clockspeeds for GTX Titan X and GTX 980 showcases just why the 50% larger GM200 GPU only leads to an average performance advantage of 35% for the GTX Titan X. While the max boost bins are both over 1.2GHz, the GTX Titan has to back off far more often to stay within its power and thermal limits. The final clockspeed difference between the two cards depends on the game in question, but we’re looking at a real-world clockspeed deficit of 50-100MHz for GTX Titan X.

 

[DownTheSky]

EDIT: delete

 

[moonbogg]

You guys think PCI-E 2.0 X16 will be enough for a cut down version of this thing? I've waited long enough and I'm dropping on those things when they hit, but I got PCI-E 2.0 X16 still.

 

[Smoblikat]

You guys think PCI-E 2.0 X16 will be enough for a cut down version of this thing? I've waited long enough and I'm dropping on those things when they hit, but I got PCI-E 2.0 X16 still.

PCI-e 2 X16 should be plenty for basically every video card out there. I watched a linus tech tips about him putting a dual GPU 6990 in an X1 slot, and it lost less than 5% performance.

 

[Headfoot]

"Losing compute" is too broad.

Compute isn't DP/FP64 throughput. They aren't the same thing. nVidia's marketing may have equated them as the same thing when the first Titan launched, but it was wrong then and it's wrong now. DP/FP64 throughput is certainly important for certain market segments, and I'm sure a good many people who would have bought the Titan X now will not because of the weak DP/FP64 performance. Compute has many, many more variables than DP throughput, like cache sizes, context switching penalties, ease-of-use for general programming, etc.

Its clearly still quite powerful at FP32. A great many graphics-compute functions like post-effect motion blur don't need FP64 because the rounding errors would come across as tiny variations in color in an effect that's intentionally smearing the picture...

Still, it's pretty clearly inconsistent with nVidia's prior marketing of what constitutes "Titan" though... The only bullet point that makes this thing "Titan" and not "Really expensive 980 Ti" is the 12 GB of memory. I see nothing else.

 

[Qwertilot]

There's also that deep learning stuff they talked about rather a lot. Implied that even 32 was too much for that, hence putting a 16 mode into Pascal and getting such a massive speed up.....

There's so much big data stuff around nowadays that this must be quite a big application area. Although promising they'll have something which will be ten times(!) faster at that sort of application in a year or so might put some of those people off buying just yet

 

[Magic Carpet]

PCI-e 2 X16 should be plenty for basically every video card out there.

Except for a few occasions such as this one:

The new Titan is likely going to be bottlenecked by the 2.0 bus in some titles. How much? Testing will show when it's ready.

 

[SirPauly]

Sorry, but blowers are dinosaurs for 250W TDP cards. For such cards, open air after-market coolers and/or AIO CLC are by far the superior solutions for getting top performance and great noise levels. Titan X cements this.

Allow the AIB's to offer choice:

http://www.inno3d.com/products_detail.php?refid=192

 

[SexyK]

I bit on one. I prefer single-card solutions over the alternative and it looks like once I get it under water I should easily be able to get +50% over the original Titan. Once I sell off my current card and water block the cost won't be too tough to swallow.

Speaking of water, I saw that EK announced their blocks last night. Looks like no availability right now though. Not sure why there aren't more options out there ready to go when the cards launch.

 

[omeds]

Reference blowers are fine, at least well engineered ones.

Shame about the price, this is one serious piece of hardware.

 

[tviceman]

Overall disapointing release, even if it is an amazing chip. I don't know if 12gb of vram cuts into efficiency that much, but it is clearly less efficient than GM204, which bucked the trend of both Fermi and Kepler big dies.

And I agree with Russian on the cooler. As amazing a blower cooler that it is, this chip is meant to be max OC'd and the blower simply cannot adequately keep up with the heat at accept able noise levels. The lower efficiency coupled with the chip'a heat leaves me wondering what Nvidia was thinking.

Nvidia left the door wide open and I hope AMD stomps them in price and performance with Fiji without a large power difference.

 

[CP5670]

There's also that deep learning stuff they talked about rather a lot. Implied that even 32 was too much for that, hence putting a 16 mode into Pascal and getting such a massive speed up.....

There's so much big data stuff around nowadays that this must be quite a big application area. Although promising they'll have something which will be ten times(!) faster at that sort of application in a year or so might put some of those people off buying just yet

Yeah, DP isn't really needed for training deep networks. I use SP for the stuff I do with them. DP is more important for scientific simulations (solving PDEs), where the problem might be inherently unstable and small roundoff errors build up. That has traditionally been the main application space for GPU compute, rather than machine learning. On the Nvidia conference blog, they're selling this as a big capability of the Titan X.

 

[96Firebird]

Allow the AIB's to offer choice:

http://www.inno3d.com/products_detail.php?refid=192

Interesting, maybe Nvidia took over selling the Titan X for a few months before AIB vendors because they are allowing aftermarket cooling this time?

 

[SirPauly]

I'm having trouble understanding, TitanX's thermals and acoustics are basically the same as the GTX 780ti.

If some desire more robust cooling for the extremer overclocker, looks like AIB's may offer more choice:

http://www.overclockers.co.uk/showproduct.php?prodid=GX-050-IN&groupid=701&catid=1914&subcat=1576

 

[AtenRa]

EK Releases GeForce GTX TITAN X Water Blocks

nice, lets see how Titan-X will OC and perform with one of those.
But you add another $100 to the already expensive GPU. And that is if you already have a custom water cooling system (rad, pump, tubes etc etc)

 

[nurturedhate]

Allow the AIB's to offer choice:

http://www.inno3d.com/products_detail.php?refid=192

That should hopefully allow this card to really stretch its legs. Will be nice to see when it is released. Also, hopefully this will also shut up some of the negative nancys going on and on about how the 390x's cooler is a affront to god, or at least tone down the rhetoric, since we are now seeing aio on an nvidia card.

 

[wand3r3r]

That should hopefully allow this card to really stretch its legs. Will be nice to see when it is released. Also, hopefully this will also shut up some of the negative nancys going on and on about how the 390x's cooler is a affront to god, or at least tone down the rhetoric, since we are now seeing aio on an nvidia card.

There are always decent aftermarket cards. It doesn't prevent reviewers from using the reference card for reviews and comparisons and people from putting on the blinders to aftermarket cards. Look at all the people blind to the 290/x aftermarket cards.

Anyways the fact that there are plenty of aftermarket cards coming points to the fact that this is 100% the 980 ti renamed to titan x to jack up the price.

 

[skipsneeky2]

Except for a few occasions such as this one:

The new Titan is likely going to be bottlenecked by the 2.0 bus in some titles. How much? Testing will show when it's ready.

I see a bottleneck already between 1.1 and 3.0,the whole review there is a couple frames difference at 900/1080p.

Doubt that means much,i thought it was funny as 1.1 has been obsolete technically since the 45nm C2D/C2Q stuff came out lol and X38.As if anyone has a q9650 and a 980 in the same system.

 

[tviceman]

I'm having trouble understanding, TitanX's thermals and acoustics are basically the same as the GTX 780ti.

If some desire more robust cooling for the extremer overclocker, looks like AIB's may offer more choice:

http://www.overclockers.co.uk/showproduct.php?prodid=GX-050-IN&groupid=701&catid=1914&subcat=1576

Basically the same, but it's coming at an expense of lower perf/w than GM204, which directly translates into lower headroom when overclocking. And whereas with Kepler, 10-15% is the normal high end for overclocking, with Maxwell it's 15-20% (with GM107 and GM204, anyways). On top of that, given that GM200's clocks are considerably lower than GM204's, there is belief (or rather, hope) that with great cooling, GM200 (Titan X) can hit 1450+mhz regularly. But as it stands now, based on several reviews, it looks like ~1250-1275mhz sustained is the overclock norm.

 

[boxleitnerb]

So now I'm in a pickle (for a 2 GPU setup):

Titan X:
(+) Much better SLI support, especially day one
(+) GPU-PhysX (yes, few games, but I like it and I play Batman and Witcher)
(+) Much better performance in CPU bottleneck. Might be critical for GTA 5 (open world)
(+) Gameworks (I don't like not being able to enable all options :'()
(+) G-Sync proven to work, 4K display at my preferred size (28") with non-TN panel at least announced
(+) DSR fully working (I concede it is somehow a moot point since it doesn't work with G-Sync+SLI and becomes rather irrelevant at 4K).

390X (includes some estimates):
(+) Slightly faster than Titan X
(+) Massive bandwidth -> good for OC scaling with core clocks
(+) Zero Core
(+) Unleashed (power target)
(+) Price
(+) CF smoothness > SLI smoothness? (I only know [H] review on this, need more sources)

I see pro's (and con's) on both sides. But especially the G-sync thing, CPU bound performance and SLI support are massively important to me. I don't know, I have to wait until Fiji release, but I fear some AMD con's won't be sorted out for some time. When did they stop releasing CF CAPs and why? If I browse GameGPU.ru for release benchmarks, many games show no CF support on launch. That is just sad. Some aren't even working today apparently.

 

[SexyK]

Assuming we throw DP compute out the window and look at the original Titan and Titan X from a purely gaming perspective, this launch actually makes a lot more sense than the original launch. This time around you are getting a fully functional big Maxwell for your $1,000 as opposed to last time around where you actually ended up with a gimped/cut down big Kepler.

With this launch, it appears highly unlikely that the subsequent 980Ti will actually be faster than the Titan X, so the price premium makes more sense as this should really be the halo product of the entire Maxwell line (barring a Titan X Black with increased stock clocks). Sure, the 980Ti will probably be several hundred dollars cheaper, but it is unlikely that it will meet/exceed the performance of the Titan X so there is some value there if you want the absolute best of the best.

This launch seems to be the culmination of nVidia's plans to create a new ultra high-end price bracket for the absolute top dog in the lineup. If you're not interested in absolute maximum performance, you can opt for the upcoming 980/980Ti and still get most of the performance for reduced cost.

 

[tviceman]

So now I'm in a pickle (for a 2 GPU setup):

Titan X:
(+) Much better SLI support, especially day one
(+) GPU-PhysX (yes, few games, but I like it and I play Batman and Witcher)
(+) Much better performance in CPU bottleneck. Might be critical for GTA 5 (open world)
(+) Gameworks (I don't like not being able to enable all options :'()
(+) G-Sync proven to work, 4K display at my preferred size (28") with non-TN panel at least announced
(+) DSR fully working (I concede it is somehow a moot point since it doesn't work with G-Sync+SLI and becomes rather irrelevant at 4K).

390X (includes some estimates):
(+) Slightly faster than Titan X
(+) Massive bandwidth -> good for OC scaling with core clocks
(+) Zero Core
(+) Unleashed (power target)
(+) Price
(+) CF smoothness > SLI smoothness? (I only know [H] review on this, need more sources)

I see pro's (and con's) on both sides. But especially the G-sync thing, CPU bound performance and SLI support are massively important to me. I don't know, I have to wait until Fiji release, but I fear some AMD con's won't be sorted out for some time. When did they stop releasing CF CAPs and why? If I browse GameGPU.ru for release benchmarks, many games show no CF support on launch. That is just sad. Some aren't even working today apparently.

Wait until 390x launches and see if Nvidia adjusts prices at all. At least then you can make a more informed decision.

 

[boxleitnerb]

Wait until 390x launches and see if Nvidia adjusts prices at all. At least then you can make a more informed decision.

That is the plan. Will be hard though if GTA V launches in April. I wish AMD would release the 390X by then.

 

[SirPauly]

Basically the same, but it's coming at an expense of lower perf/w than GM204, which directly translates into lower headroom when overclocking. And whereas with Kepler, 10-15% is the normal high end for overclocking, with Maxwell it's 15-20%. On top of that, given that GM200's clocks are considerably lower than GM204's, there is belief (or rather, hope) that with great cooling, GM200 (Titan X) can hit 1450+mhz regularly. But as it stands now, based on several reviews, it looks like ~1250-1275mhz sustained is the overclock norm.

Good post! I'm surprised by its thermals, acoustics and power output considering the sheer size and complexity of the design, being on the same node.

 

[DownTheSky]

I hate to admit it but Titan X looks like a very good GPU. My only gripe is the price. Of all the things you listed I see only 3 valid points:

Titan X:
(+) Gameworks

390X:
(+) Massive bandwidth
(+) Price

The rest will be available on both sides or are gimmicks.