[Beyond3D Forums] GP102 and GP106 coming next quarter

[Arachnotronic]

Found this interesting post on Beyond3D.

Well..given that GP102 and GP106 are basically GP104 (±) 50%..it's basically a Ctrl+C & Ctrl+V job for the engineers! (I kid of course..but compared to GP104..the additional engineering required is relatively trivial). Both GP102 and GP106 will be out next quarter according to my information..with GP107/8 being delayed a bit.

https://forum.beyond3d.com/threads/nvidia-pascal-announcement.57763/page-72#post-1926988

I don't know how credible this poster is, but I would be very excited about this if true. SLI support in games is increasingly poor (Doom for example can't even use SLI), so I am happy to see beefier single GPUs. I think we will see the Titan first to let the GTX 1080 have a bit more time in the sun, and then 1080 Ti will follow.

I'm also happy to see an NVIDIA alternative to Polaris 10 coming fairly soon. Competition is good for the consumer.

Bummer about GP107, though. The GeForce GTX 750 Ti has been my "go to" recommendation for folks trying to upgrade slim OEM PCs with lousy PSUs (no 6-pin). I hope Radeon RX 460 proves to be a good product for this market, 750 Ti is getting long in the tooth.

 

[railven]

Still no EVGA 1080 Hybrid in retail chains, so my GPU money is safe. If this gets a whiff of being true, then I can wait.

I'll schedule another day of overtime and cover the difference.

 

[Sweepr]

GP102 should be a monster. NVIDIA has quite a bit of headroom to improve on Geforce GTX 1080's performance with a larger 450-500mm² die at 250-300W.

 

[Arachnotronic]

Still no EVGA 1080 Hybrid in retail chains, so my GPU money is safe. If this gets a whiff of being true, then I can wait.

I'll schedule another day of overtime and cover the difference.

Yeah, that's what has me thinking that we won't see 1080 Ti for a while. The AIBs have barely gotten custom 1080 cards out, and putting out 1080 Ti so soon would hurt them IMO.

A $999+ Titan P or whatever isn't going to impact AIB 1080 sales too much, but it will allow NVIDIA to have something modern under the Titan branding. It's not good for the Titan branding in general for it to be at best equivalent to a 1070 and well behind the GTX 1080.

 

[Arachnotronic]

GP102 should be a monster. NVIDIA has quite a bit of headroom to improve on Geforce GTX 1080's performance with a larger 450-500mm² die at 250-300W.

Yep, it'll be interesting to see if they use HBM2 or if they stick to GDDR5X. I think they will stick to GDDR5X, as HBM2 is quite costly.

 

[Sweepr]

Yep, it'll be interesting to see if they use HBM2 or if they stick to GDDR5X. I think they will stick to GDDR5X, as HBM2 is quite costly.

If performance wouldn't change and they can get it out sooner with 12/24 GB GDDR5X, then why not.

 

[sm625]

Competition is good for the consumer.

I dont think there is going to be much competition once the 1060 launches. The most logical specs for the 1060 would be 1280 cores @ 180 mm^2. This would be one equal step down from a 1070. So there would be the same % difference between a 1070 and a 1060 as there is between a 1070 and a 1080. This would beat polaris in performance but only draw 130 watts and be cheaper to produce. Jen-Hsun has to be laughing right now.

 

[dark zero]

Yep, it'll be interesting to see if they use HBM2 or if they stick to GDDR5X. I think they will stick to GDDR5X, as HBM2 is quite costly.

As far as I know only the flagship will have the HBM2 and it comes next year.
They won't repeat the 980Ti killing the flagship scenario never again.

 

[Arachnotronic]

If performance wouldn't change and they can get it out sooner with 12/24 GB GDDR5X, then why not.

24GB of GDDR5X would just be awesome. Somehow I think that such a card would sell for >$1K though.

 

[Arachnotronic]

Jen-Hsun has to be laughing right now.

Interesting that you bring Jen-Hsun up. When somebody asked him about Polaris on NVIDIA's last earnings call, he said this:

First of all, our GPU architecture is just superior. We dedicated an enormous amount of effort to advancing our GPU architecture. I think the engineering of NVIDIA is exquisite, and our craftsmanship is really unrivaled anywhere. The scale of our company in building GPUs is the highest and the largest of any company in the world. This is what we do. This is the one job that we do. And so, it's not surprising to me that NVIDIA's GPU technology is further ahead than any time in its history.

I look forward to seeing how his claims hold up when GP106 goes up against Polaris 10 and later Vega against GP104/GP102. Exciting times for GPU buyers from both camps. We all benefit from companies working tirelessly to build excellent products. :thumbsup:

 

[Eymar]

I look forward to seeing how his claims hold up when GP106 goes up against Polaris 10. Exciting times for GPU buyers from both camps. :thumbsup:

This. Both sides are bringing their best to the table and it's quality stuff. Now what price point you want to jump in as a consumer is up to you. Each price point above $200 has good choices with my favorite being 1070 at $380. I'm hoping for GP102 to shine at 4k (above 60FPS is nice, take my money now if above 90FPS) so can ditch 1440p surround setup for 4k 32" 120hz G-Sync monitor (until this monitor exists then kind of kills that plan).

 

[Sweepr]

Interesting that you bring Jen-Hsun up. When somebody asked him about Polaris on NVIDIA's last earnings call, he said this:

NVIDIA has tailored the architecture to really take advantage of TSCM's 16nm FinFET. GP104 packs quite a punch for a 314mm² GPU.

 

[MrTeal]

A straight up 50% scaling of GP104 would be great. I do worry on price a bit, it's way too soon after GTX 1080 and availability is still too scarce to see it supplanted quite yet at a reasonable price point. They're still selling out of FE cards at $700, there's no way we're seeing a 3840 CC GPU launching in 3 months at the traditional 600-700 price point. Maybe a $1000 Titan model, which would actually make the Titan almost the same Perf/$ as GTX1080.

 

[CakeMonster]

Found this interesting post on Beyond3D.

Well..given that GP102 and GP106 are basically GP104 (±) 50%..it's basically a Ctrl+C & Ctrl+V job for the engineers! (I kid of course..but compared to GP104..the additional engineering required is relatively trivial)

Uhm, I thought the whole point was that GP104 was the "shrink" with the architectural challenges that posed, but next year's model (ie Titan and 1080Ti) would bring more architectural changes. This guy is just claiming that they add cores basically. It doesn't exactly instill much trust in his claims..

 

[Arachnotronic]

Uhm, I thought the whole point was that GP104 was the "shrink" with the architectural challenges that posed, but next year's model (ie Titan and 1080Ti) would bring more architectural changes. This guy is just claiming that they add cores basically. It doesn't exactly instill much trust in his claims..

GM200 was just GM204+50%, FWIW. I don't think it's reasonable to expect uArch changes at the SM level with GP102 from GP104.

 

[HOOfan 1]

I dont think there is going to be much competition once the 1060 launches. The most logical specs for the 1060 would be 1280 cores @ 180 mm^2. This would be one equal step down from a 1070. So there would be the same % difference between a 1070 and a 1060 as there is between a 1070 and a 1080. This would beat polaris in performance but only draw 130 watts and be cheaper to produce. Jen-Hsun has to be laughing right now.

I think there are a couple of possibilities.

1) GTX 1060 follows the performance trend of the GTX 960, GTX 760 and GTX 660 and ends up about as powerful as the GTX 970 and is priced between $200-$250

2) GTX 1060 ends up more like the GTX 660 Ti and equals the GTX 980 performance and is priced between $250-$300

 

[guskline]

NVIDIA has tailored the architecture to really take advantage of TSCM's 16nm FinFET. GP104 packs quite a punch for a 314mm² GPU.

Owning both the 980TI SC and now a GTX 1080 FE, Sweeper is correct. Very fast card.:thumbsup:

I think the GP106 is next to appear.

 

[majord]

Is he saying GP106 is half GP104? (Like GM206 was half GM204)?

If so, I don't see where people are getting the idea it will be faster than P10 by default.. It'd take some clocking to beat it if that was true.

I personally think they won't follow the same strategy as maxwell.. it left far to bigger gap between the two.


-Edit - Either that, or 1060 is a further cut down GP104.

 

[Aristotelian]

Well, I guess I could see a GP102 Titan being released this year - with the 1080Ti coming mid next year to compete with Vega. But I don't really see Nvidia doing this unless they are going to charge much more for the Titan over the 1080. I'm a member of another online forum where people say "welcome, and sorry for your wallet" as a rite of passage and...well...

 

[CakeMonster]

Well, I guess I could see a GP102 Titan being released this year - with the 1080Ti coming mid next year to compete with Vega.

I'd be willing to put money on that. Makes perfect sense and lines up with their previous release cycles.

But I don't really see Nvidia doing this unless they are going to charge much more for the Titan over the 1080

Well, we knew that already, didn't we? The financial prospects of doing that look even more favorable than 1 or 2 years ago IMO.

I have a hard time not seeing NV going this route.

What still puzzles me is if there are no decent architectural changes with GP102. I have a hard time believing that the 970/980 are basically tweakable for so many years ahead that the design would handle 100%+ performance in an ideal way. Or even get there easily.

 

[gamervivek]

He's pretty credible and he was the one who talked about the big Pascal taping out way back in July last year.

 

[Headfoot]

IMO I think we'll get 2 titans out of Pascal before volta hits. 1 will be 1080 + 50% gp102, and the other will be a full (or at least equal to the one shipping in DGX1) GP100 GPU. This is assuming Vega doesn't drop the ball. I assume a full GP100 release near the end of Pascal's lifetime (eg. ~9 months prior to first Volta drop) or as needed to take the performance crown if GP102 can't take it from Vega. If gp102 can beat Vega alone then we probably wont see a second GP100 based Titan. I bet GP102 vs GP100 will be much like 8800 Ultra g80 vs 8800 GT g92, one is probably faster by a bit, the other is lean, mean and a lot less money. This is just my speculation fwiw

 

[MrTeal]

IMO I think we'll get 2 titans out of Pascal before volta hits. 1 will be 1080 + 50% gp102, and the other will be a full (or at least equal to the one shipping in DGX1) GP100 GPU. This is assuming Vega doesn't drop the ball. I assume a full GP100 release near the end of Pascal's lifetime (eg. ~9 months prior to first Volta drop) or as needed to take the performance crown if GP102 can't take it from Vega. If gp102 can beat Vega alone then we probably wont see a second GP100 based Titan. I bet GP102 vs GP100 will be much like 8800 Ultra g80 vs 8800 GT g92, one is probably faster by a bit, the other is lean, mean and a lot less money. This is just my speculation fwiw

That would be an interesting approach; I wonder how they would actually go about it. GP100 is a natural Titan in the same vein as the original. It could have great applications as a prosumer card even if they gimp the DP from the full Tesla version.
For gaming through, I'm not sure how much faster (if any) an HBM2 GP100 Titan would be with even a full 3840 CCs vs a 384bit GDDR5X GP102 with 3840 CCs.

 

[Qwertilot]

What still puzzles me is if there are no decent architectural changes with GP102. I have a hard time believing that the 970/980 are basically tweakable for so many years ahead that the design would handle 100%+ performance in an ideal way. Or even get there easily.

They did do a lot of work to optimise Pascal It wasn't remarkable of flashy or anything but probably ultimately more useful.

They're also doing a tick/tock thing. Volta is coming at some point of course, and seems like it that might change more things. Imagine that it takes quite a while to get really working well though.

 

[Sweepr]

For gaming through, I'm not sure how much faster (if any) an HBM2 GP100 Titan would be with even a full 3840 CCs vs a 384bit GDDR5X GP102 with 3840 CCs.

Probably hardly any faster, which is why I think they will go straight to GV104 after GP102.