[Kitguru]: Nvidia delays Volta

[Mondozei]

It's all here.

Anyone still believes we'll see GP100 in consumer hands in 2016? It's going to be GP104 guys. GP100 only for enterprise, deep learning stuff. Maybe a Titan card at most. GP100 for consumers in 2017, calling it now.

 

[raghu78]

Nvidia has mastered the art of GPU releases with kepler and maxwell and have been extremely successful. With Kepler the launch strategy was GK104->GK106->GK100. With Maxwell Nvidia went one step better than Kepler and launched GM107->GM204->GM206->GM200. I expect Nvidia to follow the Maxwell strategy. I expect to see a GDDR5 based 100 - 120 sq mm GP107 sometime in Q2 2016 followed by a 300 sq mm GP104 with HBM2 in late Q3 or early Q4 2016. We will then see a 200-220 sq mm GP106 launch in Q1 2017 and finally a 500 sq mm GP100 for Tesla in Q3 2017 and Geforce in late Q4 2017. We are seeing HBM being in low supply volume in spite of being manufactured on a mature 29nm DRAM process.

http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=6757501

HBM2 will be manufactured on a bleeding edge 20nm DRAM process.

http://www2.techinsights.com/l/8892/2015-03-27/kbhjk

http://www.kitguru.net/components/g...mory-chips-opens-way-for-32gb-graphics-cards/

HBM2 supply volume in 2016 is going to be small and the real HBM2 volume will come in 2017.

 

[RussianSensation]

It's all here.

Anyone still believes we'll see GP100 in consumer hands in 2016? It's going to be GP104 guys. GP100 only for enterprise, deep learning stuff. Maybe a Titan card at most. GP100 for consumers in 2017, calling it now.

Did you read the comments under the article? Looks 100% click bait from using wccftech as a source. Pascal was a 2016 and Volta was a 2018 product as per Ananadtech's GTC 2015 blog 6 months ago:

http://www.anandtech.com/show/9088/nvidia-gtc-2015-keynote-live-blog

Also, there are rumors that NV might release the flagship GP100 first. It's hard to say right now since we have no credible rumors. In theory, I see your point since NV milked the mid-range with 670/680 and then 970/980 before releasing the true flagships (780Ti and 980Ti) but while logic tells us that they are likely to follow the same strategy again, we cannot be sure just yet.

 

[Arachnotronic]

Volta was "delayed" to 2018 a while ago.

 

[RussianSensation]

Volta was "delayed" to 2018 a while ago.

I honestly don't even see how it was delayed. NV is on a 2 year cadence:

2012 - Kepler
2014 - Maxwell
2016 - Pascal
2018 - Volta

What was delayed is Fermi and because of that, everything got pushed back 6 months, but since then NV has been going exactly according to how they intended -- bifurcating a generation into two parts, while launching a new architecture every 2 years. I guess 20/22nm mode threw some of their plans off with Maxwell but architecture wise, their 2-year cadence is on schedule based on their roadmaps since 2012.

 

[Arachnotronic]

I honestly don't even see how it was delayed. NV is on a 2 year cadence:

2012 - Kepler
2014 - Maxwell
2016 - Pascal
2018 - Volta

What was delayed is Fermi and because of that, everything got pushed back 6 months, but since then NV has been going exactly according to how they intended -- bifurcating a generation into two parts, while launching a new architecture every 2 years. I guess 20/22nm mode threw some of their plans off with Maxwell but architecture wise, their 2-year cadence is on schedule based on their roadmaps since 2012.

Volta was originally in NVIDIA's public roadmap as the follow on to Maxwell in 2016. It was pushed to 2018 and Pascal was put in its place. That is the "delay" I'm talking about

 

[AtenRa]

This is not new.

GTC 2013

GTC 2014

GTC 2015

 

[Mondozei]

Volta was supposed to be on 10 nm. I wouldn't be surprised if we end up seeing it in 2019 for consumers, given all the node delays.

Also, this still doesn't answer the question: 2017. If full fat GP100 comes out for consumers in 2016, what will NV do in 2017?

No, GP100 for consumers in 2017. Otherwise there is no plan.

 

[AtenRa]

The roadmaps are about the Architecture, not the SKUs.

 

[raghu78]

I honestly don't even see how it was delayed. NV is on a 2 year cadence:

2012 - Kepler
2014 - Maxwell
2016 - Pascal
2018 - Volta

What was delayed is Fermi and because of that, everything got pushed back 6 months, but since then NV has been going exactly according to how they intended -- bifurcating a generation into two parts, while launching a new architecture every 2 years. I guess 20/22nm mode threw some of their plans off with Maxwell but architecture wise, their 2-year cadence is on schedule based on their roadmaps since 2012.

Fermi arrived late in comparison to HD 5870. But in comparison to GTX 280 which launched in June 2008, Fermi aka GTX 480 arrived 21 months later. GTX 280 arrived 20 months after 8800 GTX which launched in Nov 2006. Throughout their history Nvidia have been launching a new generation every 18-24 months. The longest gap was 30 months from GTX 680->GTX 980 due to the delay of 16/14nm FINFET and very long life of 28nm. btw even with Maxwell the GTX 750 Ti launched in Feb 2014. So Nvidia made the architectural transition in less than 24 months. My expectation is for a GP104 launch in Sep/Oct 2016 and a GV104 launch in Sep/Oct 2018.

 

[Genx87]

This is not new.

GTC 2013

GTC 2014

GTC 2015

Looking at those pictures. It seems like Pascal is what Volta originally was to be in a tick tock cycle?

 

[dark zero]

So Volta will be avaliable on 2018...

 

[Gundark]

Looking at this pictures, Unified memory was planned for Maxwell but it was delayed. Do we have here Pascal with features that was meant to be on Maxwell and Volta delayed from original plan in 2013?

 

[lehtv]

Looking at those pictures. It seems like Pascal is what Volta originally was to be in a tick tock cycle?

NVIDIA has a tick tock cycle?

 

[Ken g6]

I don't really care what they call their next generation of GPUs, as long as it's made on a less-than-28nm process.

Although I hope it's Blaiseingly fast too.

 

[Phynaz]

What's a couple of years between friends?

Notice how the scales change on NVidia slides?

 

[RussianSensation]

Looking at this pictures, Unified memory was planned for Maxwell but it was delayed. Do we have here Pascal with features that was meant to be on Maxwell and Volta delayed from original plan in 2013?

Ya, pretty much. Maxwell was supposed to be a revolutionary architecture with unified memory, possibly Denver ARM cores but I think because 20/22nm node was borked, NV had to scrape all of that and go all in on perf/watt gaming architecture which is why Maxwell is more like Kepler refined.

They essentially just enlarged Kepler and re-organized the stream processors to make them more efficient. A lot of the core fundamentals have not been changed that much outside of just having extra units doing more work.

I am hoping Pascal will be 80-100% faster than a 980Ti because NV will have new node + hopefully a more radical new architecture + 1TB/sec memory bandwidth or 3X the current spec. The move to Pascal should be a bigger jump in theory than Kepler -> Maxwell. The last time NV only had 2 of these benefits on its side (node shrink + new architecture) they doubled the performance from GTX580 -> 780Ti.

 

[tential]

If Pascal isn't a MASSIVE leap in performance, I will be so extremely disappointed I won't know where to begin.

 

[RussianSensation]

If Pascal isn't a MASSIVE leap in performance, I will be so extremely disappointed I won't know where to begin.

There is 1 legitimate reason why it could appear as if 16nm HBM2 GPUs aren't a massive leap -- a glut of outdated, CPU limited, non-GPU intensive PS4/XB1 console ports skewing benchmarks and hiding the true potential of advanced next generation GPU architectures meant for next gen games.

HD7970 launch is the perfect example of this scenario.

At 1080P, 7970 appeared to be just 11% faster than a GTX580 at launch, and only 19% at 1440P:

It would be easy to conclude that HD7970 was a garbage release but it would have been a flawed conclusion since the GPU utilization in a lot of those games tested was probably nowhere near 100%!

We are actually already starting to see this scenario play out in many 'modern' PC games today at 1080P:

^It would be easy to conclude here that the Titan X and Nano CF are just 1.5% apart in performance and that Nano CF is only 27% faster than the 980. But I bet any $ if we look at the GPU usage data, those Nano GPUs are barely sweating (i.e., they are underutilized). Eventually we will have 16nm GPUs like Pascal that would have the performance approaching that of dual Nanos in 1 chip. However, if we continue testing them in outdated game engines/games, at low resolutions like 1080P, with hardly much of a stress, they are also going to be bottlenecked.

Turn on the GPU load (4K) and suddenly the Nanos are 38% faster than the Titan X and 82% faster than the 980!

Similarly, if you just look at 1080P comparison of 980 vs. 980Ti, you'd think hmmm..well the 980Ti is only 21% faster, but that's because it's CPU limited in many of those 1080P game tests. Throw in a GPU demanding title, and 980Ti really flexes its muscle.

That means we have to be cognizant of the fact that to show a true potential of a powerful next generation 2016-2017 GPU, we have to look at either high resolution gaming benchmarks for old/less GPU demanding games, or hope that much more GPU demanding games come out in 2016-2017 to allow Pascal to fully utilize all the extra texture, shader, ROP units, etc. Otherwise, I think it's about time that sites like TPU switch on SSAA/VSR/DSR at 1080P/1200P resolutions. It's becoming stupid to test many games at 1080P at 100-200+ fps because a lot of gamers on 60Hz monitors will start turning on SSAA/VSR/DSR if they have that much extra performance.

I hope professional reviewers start thinking about this for next generation GPUs because otherwise it's akin to buying LaFerrari and driving within the speed limit on the highway and stating that after reaching 70 mph speed limit, it's not any faster than a Honda Civic.