So is anyone actually running a GTX 1070/80?

[Stacc]

http://www.bitsandchips.it/52-engli...ds-will-probably-have-a-short-commercial-life

Well, interesting that Nvidia is pushing the GPUs so fast. And the most important: why?

My guess is that Maxwell was delayed because it was supposed to be on 20nm, not 28. Pascal was also delayed because of how long it took to get 16 nm out the door. Volta was probably originally supposed to be on 10 nm but now it will be on 16. Perhaps they are going with Intel's now defunct tick-tock strategy.

Kepler - Tick (Process)
Maxwell - Tock (architecture)
Pascal - Tick
Volta - Tock

As to why, the answer is pretty simple. Unlike all other players in the desktop/laptop GPU market, Nvidia is the only one who only makes GPUs. Both AMD and Intel also make CPUs. That means they need to push the envelope as hard as they can to justify the addition of a graphics card instead of those that come bundled on the processor. At the end of the day, it boils down to Nvidia wants to make more money.

 

[koyoot]

You still don't answer the question. Volta was supposed to be 10 nm, and suddenly now is 14 nm design? Pascal is rushed even if there is absolutely no volume in TSMC fabs, so the stock is everywhere low.

Why the hurry? There is absolutely nothing that would justify this, even milking their own customers. Because milking would mean stalling the releases as hard as you can. It is Nvidia. They like to think about themselves as Apple of GPU world.

 

[Stacc]

You still don't answer the question. Volta was supposed to be 10 nm, and suddenly now is 14 nm design?

Moving from 28 nm to 14/16 nm required moving from planar transistors to finfet, a pretty big difference. The differences in terms of GPU design between 16 to 10 isn't so big, so moving volta from 10 to 16 shouldn't be a big deal. The reason most likely has to do with 10 nm is still a ways off in the die sizes required for high performance GPUs.

Pascal is rushed even if there is absolutely no volume in TSMC fabs, so the stock is everywhere low.

This is no longer true. GTX 1070s and 1080s aren't hard to find now, and the GTX 1060 is catching up pretty quick.

Why the hurry? There is absolutely nothing that would justify this, even milking their own customers. Because milking would mean stalling the releases as hard as you can. It is Nvidia. They like to think about themselves as Apple of GPU world.

Volta isn't supposed to land until 2017. That means Pascal will have a shelf life of somewhere in the range of 7 to 19 months. That seems pretty normal. Just because they have little competition currently from AMD doesn't mean that they are going to stop designing new GPU architectures. They are trying to position their GPUs as general purpose compute machines for things like machine learning and computer vision, which means they are being threatened from not just other GPUs but CPUs and things like the Xeon Phi as well.

 

[koyoot]

Moving from 28 nm to 14/16 nm required moving from planar transistors to finfet, a pretty big difference. The differences in terms of GPU design between 16 to 10 isn't so big, so moving volta from 10 to 16 shouldn't be a big deal. The reason most likely has to do with 10 nm is still a ways off in the die sizes required for high performance GPUs.

The reason why Nvidia is moving Volta to 16 nm from 10 is... abandoning 10 nm process. Same story as Maxwell/Pascal on 20 nm. After Kepler it was supposed that Maxwell will be next gen architecture from Nvidia, and after that directly there will be... Volta. Maxwell was supposed to be 20 nm GPU. Process failed, so Nvidia had to put Pascal in between Maxwell and Volta. Same thing will happen with Volta. Expect that Volta will be slightly tuned GP100 architecture, rather than completely new architecture, because of this very reason.

We will be hearing much more about 10 nm processes in upcoming months. And I am wondering why Intel is also adding 4th refresh of their 14nm CPU lineup. Maybe for the exact same reasons.

This is no longer true. GTX 1070s and 1080s aren't hard to find now, and the GTX 1060 is catching up pretty quick.

The reason for this is high-end market got fed with GPUs, and Mainstream is completely dominated by sales of AMD GPUs. Current estimations suggest that for each one 16nm GPU Nvidia sold from May, AMD sold five(!) 14nm GPUs. And that happened in one month.

Volta isn't supposed to land until 2017. That means Pascal will have a shelf life of somewhere in the range of 7 to 19 months. That seems pretty normal. Just because they have little competition currently from AMD doesn't mean that they are going to stop designing new GPU architectures. They are trying to position their GPUs as general purpose compute machines for things like machine learning and computer vision, which means they are being threatened from not just other GPUs but CPUs and things like the Xeon Phi as well.

I know everything what you write, but this is absolutely pointless from business point of view in this case. Volta will land around April(at least first chips). So based on availability it is much lower than one full year of availability of GPU designs.

I also think that this means that GP100 architecture will go to mainstream(proper Pascal, renamed as Volta) consumers.

 

[Stacc]

The reason why Nvidia is moving Volta to 16 nm from 10 is... abandoning 10 nm process. Same story as Maxwell/Pascal on 20 nm. After Kepler it was supposed that Maxwell will be next gen architecture from Nvidia, and after that directly there will be... Volta. Maxwell was supposed to be 20 nm GPU. Process failed, so Nvidia had to put Pascal in between Maxwell and Volta. Same thing will happen with Volta. Expect that Volta will be slightly tuned GP100 architecture, rather than completely new architecture, because of this very reason.

We will be hearing much more about 10 nm processes in upcoming months. And I am wondering why Intel is also adding 4th refresh of their 14nm CPU lineup. Maybe for the exact same reasons.

Abandoning 10 nm is very different than delaying 10 nm. It looks like the transition to 10 nm is yet again going to be difficult which is why Intel is adding another interim architecture and Nvidia moved Volta to 16 nm. We don't know much about Volta, but like all GPU architectures, it will likely be iterative over the one that came before it. We also know that it is already being touted in supercomputers in 2017, which may be a hint that a HPC version is coming first.

The reason for this is high-end market got fed with GPUs, and Mainstream is completely dominated by sales of AMD GPUs. Current estimations suggest that for each one 16nm GPU Nvidia sold from May, AMD sold five(!) 14nm GPUs. And that happened in one month.

Do you have a link or any data for this? Looking at the steam hardware survey I see only the GTX 1070 and 1080 and no 1060 or RX 480.

I know everything what you write, but this is absolutely pointless from business point of view in this case. Volta will land around April(at least first chips). So based on availability it is much lower than one full year of availability of GPU designs.

Remember it wasn't so long ago that we expected new GPUs every year. Again, we don't know what markets the first Volta cards will target. If they are aiming to replace Pascal completely I still don't know why this would be "pointless." Nvidia wants to make newer, faster graphics cards to convince people to buy them over their existing cards and compared to their competition.

 

[koyoot]

Abandoning 10 nm is very different than delaying 10 nm. It looks like the transition to 10 nm is yet again going to be difficult which is why Intel is adding another interim architecture and Nvidia moved Volta to 16 nm. We don't know much about Volta, but like all GPU architectures, it will likely be iterative over the one that came before it. We also know that it is already being touted in supercomputers in 2017, which may be a hint that a HPC version is coming first.



Do you have a link or any data for this? Looking at the steam hardware survey I see only the GTX 1070 and 1080 and no 1060 or RX 480.



Remember it wasn't so long ago that we expected new GPUs every year. Again, we don't know what markets the first Volta cards will target. If they are aiming to replace Pascal completely I still don't know why this would be "pointless." Nvidia wants to make newer, faster graphics cards to convince people to buy them over their existing cards and compared to their competition.

I am not saying it is bad move to push Volta forward. All I am saying it is very strange unless there are extremely good reasons for this financial and technological, and... "competitional". Especially if you might not have enough volume in production in your fab partner.

All what I have posted is based on "whispers" from retail line. Actual numbers will come soon, at least they should come soon. However... From one Fab AMD have had 180 thousand GPUs working at the start of RX 480 sales. And in the end, the demand was so big, that all of the dies that were produced to that time, went into reference designs, and there was not enough GPUs for AIB partners.

AMD produced their GPUs in 3 fabs: Singapore, Germany, USA. And they are going to add Samsung process, because of the demand. Only two GPUs are built on GloFo/Samsung process: RX 470 and RX 480. RX 460 is built on TSMC 16 nm FF+ node. You get the picture?

 

[Stacc]

I am not saying it is bad move to push Volta forward. All I am saying it is very strange unless there are extremely good reasons for this financial and technological, and... "competitional". Especially if you might not have enough volume in production in your fab partner.

I don't understand how its "strange" for a company that makes graphics cards to release a new one a year after introducing a previous one. This is how the computer industry works. Its slowed down a bit lately but its not that far out of the ordinary. TSMC should have plenty of volume on 16nm by mid 2017 and it doesn't matter which architecture they are fabricating.

All what I have posted is based on "whispers" from retail line. Actual numbers will come soon, at least they should come soon.

I'll wait until there are actual numbers to discuss which is more successful. Remember success is about more than the number sold, its also about revenue and profit. GP 106 and Polaris 10 are roughly similar in performance but GP 106 has much fewer transistors and a smaller die. I gotta imagine Nvidia is making more money on that for each one sold than AMD with the RX 480. Unless AMD is getting a great deal from Globalfoundaries.

Remember that Apple doesn't sell the most smartphones or computers but they beat everyone in revenue and profit margin.

 

[koyoot]

When was Maxwell released? 2014. When replacements came out? 2016. When Kepler GPUs came out? 2012. When it was replaced? 2014. It goes one by one architecture from Nvidia.

Pascal might not live more than 10 months! That is strange.

P.S. Why do you think I am saying that one or another company was successful?

 

[Stacc]

When was Maxwell released? 2014. When replacements came out? 2016. When Kepler GPUs came out? 2012. When it was replaced? 2014. It goes one by one architecture from Nvidia.

Pascal might not live more than 10 months! That is strange.

Nvidia is increasing their pace of their development so that their GPUs can compete in an increasing number of compute and graphics tasks. I don't really understand how that is "strange."

P.S. Why do you think I am saying that one or another company was successful?

I don't know what you are getting at with this statement.

 

[Asgorath]

When was Maxwell released? 2014. When replacements came out? 2016. When Kepler GPUs came out? 2012. When it was replaced? 2014. It goes one by one architecture from Nvidia.

Pascal might not live more than 10 months! That is strange.

P.S. Why do you think I am saying that one or another company was successful?

I'd like to see some better sources than "retail whispers that I heard" before I believe any of this, yeah.

 

[koyoot]

I'd like to see some better sources than "retail whispers that I heard" before I believe any of this, yeah.

For example Gibbo from oc.uk, site that has sold highest number of 16/14 nm GPUs in whole Britain. He has provided us with numbers, both for Pascal and Polaris. Pascal happens to be best selling high End GPU, Polaris - best selling GPU overall.

He have said that they have shipped 2000 GPUs in first month of GTX 1080 sales.

They also sold 2000 RX 480 in first 48h of sales.

But lets not get bogged down in brand war. There is more important factor to think about.

Nvidia is increasing their pace of their development so that their GPUs can compete in an increasing number of compute and graphics tasks. I don't really understand how that is "strange."



I don't know what you are getting at with this statement.

Simple as it can be. How come they claimed that they spend billions of dollars on R&D of new GPUs and then they immediately jump into next big thing - Volta. From business point of view as a whole - it is much more strange move than you think.

It is not bad. But I wonder what that does imply in big scheme of things. And no, it is not usual thing to push GPUs currently so hard, and push another architecture forward on closer node. Could be abandonment of 10 nm process, but is it only thing? That is what is interesting. I guess we only will know after a year or so.

 

[Stacc]

I'd like to see some better sources than "retail whispers that I heard" before I believe any of this, yeah.

For example Gibbo from oc.uk, site that has sold highest number of 16/14 nm GPUs in whole Britain. He has provided us with numbers, both for Pascal and Polaris. Pascal happens to be best selling high End GPU, Polaris - best selling GPU overall.

He have said that they have shipped 2000 GPUs in first month of GTX 1080 sales.

They also sold 2000 RX 480 in first 48h of sales.

But lets not get bogged down in brand war. There is more important factor to think about.

Yeah, I'll wait until we get better numbers than a random post on a forum. Even if its true its a very limited set of data.

Simple as it can be. How come they claimed that they spend billions of dollars on R&D of new GPUs and then they immediately jump into next big thing - Volta. From business point of view as a whole - it is much more strange move than you think.

Its not like money spent on Pascal development is wasted. All the improvements in pascal get rolled into Volta and then Volta will have its own set of improvements. Nvidia has to stay on top because their only business is GPUs. AMD can afford to release a mid-range GPU and sit on it for 9 months because that mid range GPU will go into millions of consoles. Nvidia does not have that luxury.

 

[koyoot]

Its not like money spent on Pascal development is wasted. All the improvements in pascal get rolled into Volta and then Volta will have its own set of improvements. Nvidia has to stay on top because their only business is GPUs. AMD can afford to release a mid-range GPU and sit on it for 9 months because that mid range GPU will go into millions of consoles. Nvidia does not have that luxury.

Sorry, but I do not buy this argument for a second .

Polaris is not console GPU, and it will not go into consoles. Xbox One S has slightly tuned previous version of APU inside, and ported to 16 nm.

 

[Stacc]

Polaris is not console GPU, and it will not go into consoles. Xbox One S has slightly tuned previous version of APU inside, and ported to 16 nm.

Sony has already announced they are working on a console for release next year that will contain an enhanced GPU. Anyone want to take bets which GPU that turns out to be? I don't think its coincidental that Sony is coming out with a VR headset and Polaris 10 was designed to meet the minimum requirements for VR.

 

[Asgorath]

For example Gibbo from oc.uk, site that has sold highest number of 16/14 nm GPUs in whole Britain. He has provided us with numbers, both for Pascal and Polaris. Pascal happens to be best selling high End GPU, Polaris - best selling GPU overall.

He have said that they have shipped 2000 GPUs in first month of GTX 1080 sales.

They also sold 2000 RX 480 in first 48h of sales.

But lets not get bogged down in brand war. There is more important factor to think about.

I'm not talking about your sales figures, I'm talking about Volta coming 10 months after Pascal. I would also take your sales figures with a massive grain of salt. Who is the "us" you're talking about when you say "provided us with numbers", exactly?

 

[koyoot]

I'm not talking about your sales figures, I'm talking about Volta coming 10 months after Pascal. I would also take your sales figures with a massive grain of salt. Who is the "us" you're talking about when you say "provided us with numbers", exactly?

Readers of overclockers.co.uk forum. Gibbo is the owner and the admin of this site. And they sell GPUs. They have sold highest number of GPUs over United Kingdom(14 and 16nm ones).

 

[Asgorath]

Readers of overclockers.co.uk forum. Gibbo is the owner and the admin of this site. And they sell GPUs. They have sold highest number of GPUs over United Kingdom(14 and 16nm ones).

And he knows NVIDIA's release plans for the next couple of years? They haven't even finished rolling out all their Pascal GPUs, so I find it hard to believe that someone like that would be intimately familiar with the exact roadmap from NVIDIA.

 

[koyoot]

No he doesn't. I am talking about sales numbers of current 14/16nm crop of GPUs. Volta pushed forward is media information. You should do your research properly. I can be found on number of sites, already.

https://www.techpowerup.com/224413/nvidia-accelerates-volta-to-may-2017

 

[Asgorath]

No he doesn't. I am talking about sales numbers of current 14/16nm crop of GPUs. Volta pushed forward is media information. You should do your research properly. I can be found on number of sites, already.

https://www.techpowerup.com/224413/nvidia-accelerates-volta-to-may-2017

I think you're reading way, way too much into this:

Following the surprise TITAN X Pascal launch slated for 2nd August, it looks like NVIDIA product development cycle is running on steroids, with reports emerging of the company accelerating its next-generation "Volta" architecture debut to May 2017, along the sidelines of next year's GTC. The architecture was originally scheduled to make its debut in 2018.

Here, "debut" probably means the first time they start talking about the details of the architecture, not products go on sale. Same thing happened with Pascal, they were talking about the architecture last year but it actually shipped this year. For example, here's an article that talks about Pascal from GTC 2015:

http://wccftech.com/nvidia-pascal-gpu-gtc-2015/

So yeah, Pascal products aren't going to have a 10-month lifetime, they're just going to start sharing details of their next-gen architecture sooner than expected.

 

[Stacc]

We haven't talked about it much on this thread but reviews are trickling out on the new Titan X thats based on Pascal. This is a monster of a card. It has something like 11 TFLOPS of performance and destroys any other GPU out there (at least in gaming). I haven't found any reviews that dig into compute performance, but that should be impressive too.

 

[Squuiid]

Well, there appears to be a stutter/DPC latency issue with Pascal cards. NVIDIA have tried to fix with a 'hotfix' but it does nothing to improve the situation, nor does the newly released 369.09.
I'm returning my 1080 tomorrow.

50 pages on NVIDIA forum so far... https://forums.geforce.com/default/...s/gtx-1080-high-dpc-latency-and-stuttering/1/

1151 posts on Overclock... http://www.overclock.net/t/1605618/ongoing-pascal-latency-problems-hotfix-doesnt-work-for-everyone

 

[koyoot]

http://semiaccurate.com/2016/06/24/serious-manufacturing-problem-hits-nvidia-consumer-gpus/

This is not software related. This is hardware problem.

 

[Squuiid]

http://semiaccurate.com/2016/06/24/serious-manufacturing-problem-hits-nvidia-consumer-gpus/

This is not software related. This is hardware problem.

Yep, I'm starting to suspect as much which is why I'm not hesitating to return the card after only a few weeks of ownership.
Pascal appears to be very broken.

 

[Stacc]

Hmm, I've not noticed any problems on my GTX 1080.

 

[koyoot]

All I can tell you is that it is related with memory controller and memory itself. Nvidia used GDDR5X pre-massproduction samples of GDDR5X to get enough of them, out of the garage(similar thing with HBM2 on Pascal GP100 that were delivered at end of June 2016).

So if you did not seen any problems with your GTX 1080, you have not seen them yet.