Meet the Pascal family: Nvidia reportedly working on GP102-based GTX Titan and GTX 1080 Ti

Scorpus

Posts: 2,156   +238
Staff member

Even though we've only just seen the launch of Nvidia's GeForce GTX 1080, new rumors have started to appear that detail two faster graphics cards based on upcoming Pascal GP102 silicon: a new GeForce GTX Titan, and a GeForce GTX 1080 Ti.

A report from ChipHell, via The Tech Report, has allegedly detailed these two new graphics cards along with a more modest GTX 1060. The specification table they posted, seen below, includes some interesting tidbits if they turn out to be accurate.

A fully unlocked version of GP102 appears to be included with the Titan, packing 3,840 CUDA cores, up from 2,560 in the GTX 1080. This is even more CUDA cores than the GP100-based Tesla P100 (which has 3,584 cores), although GP102 seems to be limited to 1/32 FP64 performance rather than the 1/2 seen in the P100.

The GTX 1080 Ti would see the CUDA core count cut to 3,456 along with a reduction in TMUs. Both the 1080 Ti and Titan are listed as having the same clock speeds and memory systems (GDDR5X), along with the same TDP of 250W, although for high-compute environments, the Titan packs a huge 24 GB of VRAM as opposed to the 1080 Ti's 12 GB.

As for the GTX 1060, ChipHell has this card listed as using an even further cut down variant of GP104 with just 1,280 CUDA cores. For reference, the fully unlocked GP104 used in the GTX 1080 has 2,560 cores, while the cut down variant in the GTX 1070 features 1,920.

The GTX 1060 is also listed has having slightly lower clock speeds and reduced memory bandwidth, along with a lower TDP of just 110W.

At this early stage it's not clear whether these specifications will turn out to be true. Nvidia hasn't even released the GTX 1070 yet, so it'll be a while before we hear any details on the GTX 1080 Ti or new GTX Titan.

Permalink to story.

 
They've more than doubled the amount of RAM the stock GTX 960 has from 2 GB to 6 GB but haven't even doubled the memory bandwidth or bit size. Whether this is a bottleneck or not is going to depend on if the rest of the GPU can keep up and if that memory compression is any good.

One thing is for certain, having 3 cutdown versions of the same die means Nvidia has allot of imperfections in their yields, if they didn't they wouldn't be able to produce so many cutdown cards. The flipside of this is going to be that Nvidia is only going to be able to get a small amount of cards per wafer. The 1060's die is significantly larger than that of the GTX 960 (227mm2).
 
Looks like there's huge gaps in performance between then 1060 - 1070 and 10700 - 1080 hopefully to be filled with ti versions and even more hopefully with G5X.
 
If nVidia makes the new Titan without HBM2, nobody in the gaming community will buy it. 12GB in the current Titan X makes a small difference from 8GB of the Ti cards. But there won't be any difference for gamers to go over over 12GB of VRAM.
 
Chiphell....ROFL. Someone from a Chinese forum makes a table from the guesses of the 3DC forum, and it's now passed off as fact. Well done internet.
If nVidia makes the new Titan without HBM2, nobody in the gaming community will buy it.
Not sure how you arrive at that conclusion. Unless the issue of localized heating from the interposer package, or the added sensitivity that stifled Fiji overclocking have been overcome, then the only real issue is if the GPU is bandwidth starved.
Maxwell and GP 104 have not been demonstrated to have the issue in gaming workloads. The issue has been texture fillrate.
Also, there is small but highly visible consumer base that simply buy the fastest card - for epeen and benchmarking. If the GP 102-based Titan fits the criteria, why wouldn't the trend continue?
I'll let @dividebyzero explain that for the 100th time. I'll give you a hint though "compute".
Compute, but very likely rendering and visualization. A12GB framebuffer is barely adequate for rendering in 4K. 5K and more importantly, 8K, need much higher framebuffers to contain a frame. The prime reason that the M6000 is now available as a 24GB card.
For computing purposes there is the new Tesla M10.
Nope. Different class of card entirely. Four GM 107 GPUs with 8GB each.
Plus there will be a new Pascal-based Tesla, with HBM2 memory. Now that will be a kick-a$$ computational monster. Titan X - not so much.
The P100 costs around $10,000 each. Hardly the same market as semi-pro hobbyist renderers and the majority of businesses that incorporate CAD for advertising, marketing, small industry etc.
 
Last edited:
If nVidia makes the new Titan without HBM2, nobody in the gaming community will buy it. 12GB in the current Titan X makes a small difference from 8GB of the Ti cards. But there won't be any difference for gamers to go over over 12GB of VRAM.
You're making the big mistake of underestimating some peoples stupidity, and some of those numb nuts have deep pockets nVidia won't make it if they can't sell it. Take this article at it's face value.
 
Smart people might want to look for a cheaper 8GB OEM variant for the 1080 TI. 12GB does nothing but inflate the price.
LOL. How do you suppose you get 8GB with a 384-bit interface? Smart people aren't going to be looking for 8GB as an option.

384-bit means 6*64-bit dual channel IMC's multiplied by 2 * 8Gbit memory chips per controller. That means the options are 12GB (or 24GB in clamshell mode)
 
What would be the reason if they don't make gp 102s without hbm2? hbm2 production had problems, didn't catch up with the chips' mass fabrication? or maybe nvidia wasn't impressed by the performance (or cost or power efficiency?) hbm2 provided vs G5X?
 
LOL. How do you suppose you get 8GB with a 384-bit interface? Smart people aren't going to be looking for 8GB as an option.

384-bit means 6*64-bit dual channel IMC's multiplied by 2 * 8Gbit memory chips per controller. That means the options are 12GB (or 24GB in clamshell mode)
nobody said anything about keeping the 384-bit interface. a slightly higher clocked GDDR5X memory should be enough to get close to 400GB/s. the difference in performance at 4K should be only 1 or 2 FPS. (this might also drop the Pin requirements from 6+8 to just 8, but I'm not sure because of the larger GPU))
 
Last edited:
And nobody was surprised.
Smart people might want to look for a cheaper 8GB OEM variant for the 1080 TI. 12GB does nothing but inflate the price.
And how often does a GPU comes with LESS memory? More is common, but less?

And besides, with a GPU as powerful as the 1080ti, "cheap" has no place here. You dont cheap out on a halo product. Considering the settings dual 1080tis could push, you'd need that much vram.
 
Last edited:
And nobody was surprised.

And how often does a GPU comes with LESS memory? More is common, but less?

And besides, with a GPU as powerful as the 1080ti, "cheap" has no place here. You dont cheap out on a halo product. Considering the settings dual 1080tis could push, you'd need that much vram.
like it or not min/maxing a PC build will always involve a budget. not everybody has the money to buy the best components. even 50$ can help a lot when building a PC.
 
If nVidia makes the new Titan without HBM2, nobody in the gaming community will buy it. 12GB in the current Titan X makes a small difference from 8GB of the Ti cards. But there won't be any difference for gamers to go over over 12GB of VRAM.

HBM2 has yet to hit production, that is why they go with the fasted currently on the market with GDDRX5.. I guess what it will sell regardless
 
If no one in the gaming community purchased a Titan, the Titan would still sell.

And who will be buying it?

just about anyone in a large scale production environment.. that can be animation, 3d modelling, game dev, even video encoding outside of a manufacturing environment there are a lot of professionals that will take the high end consumer model over the business line (qaudro/fire pro) as the consumer line is usually lower on power consumption and higher on the memory, but I dont agree with the original statement the titan will sell fine to the gaming community, there is always someone that needs the latest and greatest in the custom build and real life capabilities.. maybe three screen 4k on a single card? w/good settings frame rates
 
Nvidia is panicking with the announcement that AMD's Vega GPU was pulled from 2017 to Q4 2016. Seems dropping HBM2 was the only way Nvidia could match AMD's timeline now. It'll be interesting to see how this year ends up.
 
I'll let @dividebyzero explain that for the 100th time. I'll give you a hint though "compute".

For computing purposes there is the new Tesla M10. Plus there will be a new Pascal-based Tesla, with HBM2 memory. Now that will be a kick-a$$ computational monster. Titan X - not so much.

The M10 is a monster for FP64 calculations. Many in the market, like myself, might buy a Titan variant because there are few BOINC projects that require FP64. In particular, the one project that gives the most credits for work done, GPUGrid, is FP32. Lots of people who run GPUGrid run Titan X and 980Ti right now and they crunch a lot of work units because the project is FP32. Milkyway@home is one of the few projects that requires FP64, however, credits given for the amount of work is nowhere near the credit given by GPUGrid. Pascal in any consumer flavor will likely pack a lot of compute power even though it is only in the FP32 realm, and at 1/32 for FP64, I expect that it will do better at FP64 than previous generation cards, however, I will wait for someone to run bench marks before I make my decision on what to buy.

As I see it, those with deep pockets may run the M10, but it is likely targeted at the supercomputer market more so that the consumer or prosumer market.

nVidia sells variants like these, at least as I see it anyway, because people buy them for whatever reason and they make a lot of money on them.
 
like it or not min/maxing a PC build will always involve a budget. not everybody has the money to buy the best components. even 50$ can help a lot when building a PC.

The option your describing is a GTX 1080, why would Nvidia bother releasing a lower end GTX 1080ti if they have a card more or less in that segment already? I think they're doing a good jog catering to many demographics across the price range already. When building a $2000+ PC a difference of $50 is nothing, I quoted many PCs in my days working at a local PC shop and generally if people had the option of saving a little or getting better performance they went with the better performance. In the $1000 range fine $50 could make or break the deal to some, but it's still not a considerable enough margin that people would be unwilling to dish out for the better product.
 
People have to be crazier than s*** to buy anything for dead-end PC gaming. I kicked PC to the curb a year and a half ago for PS4 and haven't looked back. I might even get the new PS4.5 for 4k gaming yes 4k on PS4. Wake up before you are all dead broke!

The reasons I left PC was #1 no gamers left on PC because of the lack of competition. One guy running a high end card running 100+fps and the others running a $100 card at 15 fps. When I gamed I always had the best card or 2nd best and SLAUGHTERED everyone. PS4 everyone plays on the same page in fps no one has an advantage with hardware.

Dumping PC for PS4 was the best thing I've ever done for my gaming.
 
Back