Nvidia GeForce GTX 1080 Ti Arrives late March

No, it was supposed to be Pascal. The blocks are 2 years apart, the Maxwell was on 2014, Volta was on 2016. They split Volta into two because there was no way they were doing HBM2, 16nm transition + all the Pascal adds + all the things Volta add in the same release.

And as I already stated, Volta isn't 10nm.. AMD/Nvidia are both skipping 10nm. Volta is rumored to be on the 16nm refresh TSMC is calling "12nm".

 

Pascal wasn't any kind of serious architectural departure, and they were already doing 16nm and HBM 2 with it (see GP100). Whatever the issue is with Volta, it sure isn't that. Also don't tell me that in that 2014 roadmap they didn't know that they had to do 16nm+stacked memory, because they clearly did.

Whatever the issue is with Volta, it isn't the transition to 16nm, nor the memory. It was supposed to be that product for NVIDIA from the start.

 

No. Road map images change, their own public statements are contrary. It is what it is.

2014 news release:

The systems will also feature NVIDIA's future generation GPU architecture, Volta™, which will deliver considerably higher performance than the company's current Maxwell™ architecture and subsequent Pascal™ design. Delivering significantly higher levels of computational performance than anything available today, NVIDIA GPUs will provide Summit and Sierra with more than 90 percent of the peak floating point processing capability. - See more at: http://nvidianews.nvidia.com/news/u...puters-for-national-labs#sthash.usPTUfT5.dpuf

 

Whatever the issue, whatever the roadmap... it's working for them.
That's the issue. There is no issue.

Nvidia keeps executing like a ****ing machine, Quarter after quarter after quarter, while still being deep inside bleak PC.

 

They've got a decent product and the competitor does too. I think I am like most. I've developed a comfort zone using Nvidia. I've used them long enough that I'm aware of the quirks and most of the tricks to maximize performance. AMD users are the same. Each has products that cover the range of consumers form entry level to extreme enthusiasts. So, other than nibbling at the enthusiasts crowd margins, what is the incentive for switching with comparable products?

 

Well, if I could be convinced that NVIDIA was presenting something at $500 that I could hold for at least three years, I would seriously consider it. I don't mind the quirks, some change would be refreshing.

But I somehow I get the feeling that they will try to convince me to give ~$750 for something with GDDR5x and Pascal in it, while I am fairly certain that due to Scorpio, I could comfortably keep any big Vega product for at least four years. NVIDIA does well with their efficiency, but in the long run it's always the beefier hardware that wins. My card is on the same ballpark with a 780Ti for some newer games, which is ridiculous.

 

Isn't that more because of AMD's poor DX11 driver? I'm not sure what that old argument has to do with anything and I very much doubt that paradigm will repeat itself...unless you're hinting at future incompetence in AMD drivers
I was reading here on Guru3D not too long ago that Nvidia was the one using 'brute force' approach versus AMD's more elegant...I guess it all depends upon which approach you prefer.

 

No, I agree about the DX11 driver. It's more than that actually, it's the architecture itself that was really made for a paradigm used today, instead of when it launched. On the other hand if it's 12.5Tflop and HBM 2 vs 10Tflop with GDDR5x, and I'm planning to keep it for longer, I believe there will be no competition.

AMD GPUs have had issues with geometry, but even Polaris has solved most of them. If any of the Vega architecture previews hold any weight to them, this will be most likely something like an ultra-efficient Fury X with 2x better frontend and 8GB of memory at 1.5GHz. Even if the 1080 is a bit faster on launch (and I don't really believe it will be), there is no comparison of what I would choose.

I secretly hope that both Volta and Vega launch close to each other, with comparable performance. And for once I hope NVIDIA goes the beefier route, the way they're doing lately it seems that you can't really stay on the same level with the equivalent AMD product past the two year line.

 

Okay...one example for AMD getting driver enhancements over time for an architecture resulting in significant gains, does not a trend make. Maybe it does validate the need for investing in expertise in firmware and drivers as I do not see them repeating the same mistake of leaving so much untapped potential unused. Imagine if AMD had that performance earlier! How much money did they lose? The rest is speculation. There's nothing wrong with that. But future speculation doesn't create a trend.

 

Only if AMD screws up royally. Consumer Volta being yearish away. It will leapfrog Pascal/Vega. +70% SGEMM/DGEMM. With not as much BW increase.

 

What?

What do you mean by already doing it? GP100 is Pascal. They made the decision to split Volta because it was too much for one release. In the slide it even shows the marquee feature for Volta being stacked ram.. which obviously is in Pascal. I should also point out that when this slide was released, Nvidia was still planning to use Hybrid Cube Memory for it's products. So while I think NVidia knew about 16nm - big changes do occur pretty rapidly.

Also Pascal has more architectural changes then just 16nm and HBM2. On GP100 it has packed math and NVLink. They optimized critical path and employed custom cell libraries. They also updated the display block for the VR stuff + playready, HDR, 10bit, etc. They improved memory compression and pre-emption + dynamic load balancing. Plus you have other unlisted changes which Nvidia does from time to time.

Since 2015, Volta has been slated for a 2017/2018 release. You keep saying there is a problem with it, but there has been no indication of anything.. not even rumors. You're basing it on it being delayed 3 years ago and again most people say it wasn't delayed it was just split over two releases.

 

PR does his homework on AMD based things but he has a habit of half-assing anything to do with Nvidia.

I still love you PR.

 

Well like 90% of the Volta being split stuff is like speculation. We know stacked DRAM was the main feature of it and we know Pascal got that. We know they moved Volta back and we know that both companies generally like playing with architecture less on new nodes (too much that can go wrong). On top of that, we don't even know if the names Nvidia is using are "anchored" to the architectures. They could have simply liked the name Volta for something else in the pipeline and swapped the names.

But there has been zero indication of Volta (as in the next architecture) having any problems or issues. Jen's said they'd be sampling Xavier (Volta based) units by Q4 this year and we know that the two super computers with Volta were supposed to be finished by this year. It wouldn't surprise if me if they got delayed to Q1 though as it feels like everything gets delayed now in days.

I personally don't think we'll be seeing consumer based Volta units till Q1 anyway. Past few new arch releases have been 1 year from the previous arch's Ti. I think they'll probably go for a similar time frame here. What we've seen from Vega looks good - but I don't think Nvidia will need a refresh to compete with it, even if the Ti is slightly slower in newer titles - they could ride that for a year easily. Especially since the 1080/1070 locked a sizable chunk of the market up.

 

I think you two should re read what he said.

I'm pretty sure he said that GP100(pascal) already has HBM2 so volta delay is not because of that.

 

No one said Volta was delayed because of HBM2. He quoted me and what I said was that it was speculated that Volta was split into two releases, Pascal/Volta because Nvidia couldn't bring HBM2 and a Die Shrink + all the architecture features of both into one release. Both companies generally tend to keep their node shrink releases small in terms of architecture. So instead Nvidia chose to do the HBM2/Die Shrink + some of the changes with Pascal - then follow up with a larger architecture change with Volta. Which is also where all the rumors of Volta being a radically different architecture started.

I don't know if I agree with that or not but it seems reasonable given Volta's original "stacked dram" feature, which is now part of Pascal.

Either way it doesn't matter, that delay happened nearly 3 years ago - since then Volta has been on track for 2017 and there has been zero issues reported about it. Not even from Charlie Demerjian who loves reporting on Nvidia's delayed architectures.

 

There are driver enhancements, and the way that games are made. Games made today are closer to GCN, by the fact of the console generation itself. This will continue for at least the next four years (Scorpio this year plus an extra three for the console itself at a minimum). It's a combination of factors that lead to this. Vega as an architecture will come about in a much more "GCN friendly" environment, so I don't believe that it will take it as much time for the differences to show. I also honestly believe that GCN is harder to optimize for the DX11 driver, but it's probably being addressed with Vega.

What I mean is that obviously NVIDIA can do new architecture + new node + new memory. They did that with GP100. There is no reason that Volta wasn't that, except the fact that Volta is once more delayed. Only this time we most likely know it wasn't either the 16nm transition, neither the HBM that was the reason.

I'm sure. But we both know it was closer to Fiji --> Polaris, than it was to Kepler --> Maxwell. Pascal is an excellent "play safe" product, where you get what already works great, improve a few parts, add some others and you basically get a revision. I don't want to be pedantic, as we both know that Volta has been rumored to be much more than that. Just think about how long we've been hearing the codename, and how many times it has been delayed.

It was late three years ago. Already. Mid-2013 there was no wording for "two phases" or anything else. That was Anandtech, and here's the same link from NVIDIA's blog itself. There is no mention of anything between Maxwell and Volta. Here it is from the leather-jacket man himself.

The "split" option makes no sense. Pascal is basically a Maxwell derivative, even by your own measure (a familiar architecture for a new node). I have never heard a tech company swapping architecture names. If they did, there would be no Volta down the line. Pascal would be Volta, and what we now know as Volta would be called something else.

Even if they sample it by Q4, it's already more than three years late according to their initial planning. Pascal was great for NVIDIA because more or less covered for Volta. That doesn't make Volta any less delayed (by NVIDIA's own estimates). It is already delayed.

My bet is end of Q1 2018, if that. If Vega ends up in mediocrity, I can see them delaying it even more and wait for 10nm. The key component to all this, imo, is Vega. If AMD has indeed managed to fight most of their efficiency bottlenecks, then it's a done deal until Volta. I'm not sure that they will manage that in the timeframe for the original Vega reviews though, although this time they took the interesting step of giving hardware to the enterprise/HPC crowd earlier (from March and on), and try to give time to their guys to go on with the driver.

From posts in the Linux kernel mailing list from the AMD driver guys, they all mention how all of the driver team is doing nothing else but optimizing for Vega, so there is a bit of hope there. Volta at 10nm and whatever AMD makes next at 7nm, means that they probably will never directly compete again, generation wise.

 

In the video you linked he says that stacked vram was the main feature of Volta, which is obviously in Pascal. So clearly the Volta coming out in 2018, is not the same Volta as he was talking about in 2014. Which is the whole point - the features planned for the 2016 version of Volta were split between what we now know is Pascal and 2018 Volta.

To make it more clear - In 2014, Nvidia stated they'd have an architecture called "Volta" which would release in 2016 - the main feature of this volta was stacked VRAM. In 2016 Nvidia released an architecture called Pascal which featured stacked VRAM. So obviously The Volta of 2014 either became Pascal or was split between Pascal and 2017/18 Volta.

Also you keep saying it's 3 years late (actually in this post you said it was late 3 years ago lol) from the initial release, but the initial release was 2016. Those blocks are 2 years apart.

And again, Nvidia is rumored to be skipping 10nm. Volta is rumored for 12nm, which is a 16nm refresh.

http://technoaisle.com/nvidia-reportedly-skipping-10nm-lithography-volta/

http://hothardware.com/news/nvidia-12nm-finfet-volta-gpu-architecture-replacing-pascal-2017

 

Alright...lets not do another thread of this same old same old....nothing has changed in the arguments in several thousand posts.

 

So we're back to somebody obviously neither owning nor interested in buying a Nvidia card throwing in a few lines and highjacking threads? Got to love guru forum these days

 

Well. Thing about that 2014 roadmap is:
- Pascal is already in DP performance per Watt range planned for Volta (16nm manufacturing saw to that)
- Pascal's Tesla P100 delivered on Volta's planned Stacked DRAM (IMC redesign)

So, with above fulfilled nVidia could easily name those GPUs as Volta. And as such, I would say that Volta got renamed to Pascal. Because Volta will be used for GPUs which have some bigger technological changes under hood since changes from Maxwell to actual Pascal are not that magical.

(Yes, I say it in contrast of Huang's statement about doing few miracles with Pascal.)
- - - -
By that I mean that you guys above are right in a way. On one side nVidia delivered what was planned for Volta. But at same time they did butt in Pascal name for no other reason that to keep name Volta for next iteration.
(Volta has been delivered, but is still delayed to 2018.)

On that note, Originally named Volta GPUs (2014 roadmap) should have deliver ~20 DP GFLOPs/Watt and come in 2016.
Then there is 2015 roadmap where Pascal "absorbs" some goals for Volta (moved to 2018).
And then another 2015 roadmap which mentions Volta's plans as 72 SGEMM/Watt, which I would roughly translate to 36 DP GFLOPs/Watt because that's where nVidia can introduce next grandiose jump in performance.

Spoiler

TL;DR: nVidia makes terrible and inconsistent roadmaps. It is like they change their mind about goals every 6 months. One day roadmap goals are DP/Watt, other day it is memory bandwidth, then it is single precision per watt.