Nvidia Volta Speculation Thread

[ImSpartacus]

Why do you think Vega should have launched together with Pascal?

I think he's disappointed that GP104 and GP102 went unanswered for a roughly a year, hence the comment about a GP102-based 1080 Ti "neutralizing" Vega 10.

Recall that consumer GP102 hit the market in August 2016 in the form of the Titan X. If Vega had released in late 2016 as originally projected, then it would've been in a pretty decent spot to at least put up a fight in the high end. But it seems like Vega 10 will drop in late spring-ish, having given Nvidia unchecked ownership of the highend for quite a while.

Therefore, I, too, would be surprised to see consumer GV104 before 2018. Nvidia just doesn't need to. If pressured, Nvidia could easily release a fully enabled GP102 (maybe even with faster 11-12Gbps GDDR5X) in order to finish out 2017 on top (and they've shown that halo performance crown is important to them).

 

[lanek]

The general product roadmap really should not be used as a literal interpretation for actual date of launch.
While the general product roadmap may seem to infer 2018, the specific presentations had it down as 2017, in fact I have a very early 2015 Nvidia HPC presentation that showed Pascal as 2016 and Volta as 2017 along with later presentations.
In same way in those presentations show the IBM/Nvidia Summit (mostly completed by 2017) is to be replaced in 2022 looking for ~20MW supercomputer.

If using Pascal as the 1st iteration of the strategy:
January 2016 announce Tegra Drive PX2
April 2016 announce Pascal and P100
June 2016 GP104 starts to become more available
August 2016 GP102 Titan starts to become available
Mid-late Q3 2016 DGX-1 available accessible to more businesses (we were offered this option with delivery being 1 week).
Late Q3 2016 GP102 Tesla/Quadro starts to become available
Late Q3 Drive PX2 goes into sampling with manufacturers
Early-mid Q4 2016 P100 available in certified solutions own nodes same config as DGX-1 (we were offered this option as well with delivery 2-3 days).
early-mid Q1 2017 GP100 launched as a Quadro product

These dates are pretty important because:
January 2017 announce Tegra Xavier to replace PX2
~May 2017 more indicators that V100 will be announced (GTC this year a month later than laster year that was April)
~Q3 2017 Nvidia has stated this is when Tegra Xavier is going into sampling with manufacturers - same period as the Drive PX2.
~End 2017 Summit supercomputer is meant to be mostly complete, going live is somepoint 2018 but has to take into consideration thorough testing/training users and importantly system admins/coding changes-porting/optimisation/etc.
~Q1 2018 hands-on workshops with it to start training users.
I think one of the other supercomputer projects associated with Summit is meant to be complete by Q1 2018, the 3rd associated project not sure.
This is important as Summit requires over 20,000 V100 dGPU (6xper node and 3,500 nodes but number may had increased as scope criteria may had been upped), and the two other supercomputers have another 20,000 V100 between them.
So that is a heck of a lot of GV100 accelerators just for 3 supercomputer contracts let alone additional orders, consider that the GP104 was meant to have launched according to some with ~30,000 GTX1080.

The only unknown is whether Nvidia will launch a consumer card like 2016 with the GTX1080, I tend to think they will launch very late Q4 2017 a GV104 consumer product.
It is not in their interest to have the gap too large between the 1st Tesla card on a new architecture and the consumer card otherwise it is a headache for the next architecture product cycle, possibly one reason Nvidia accelerated the launch of certain cards such as Pascal Titan.

Anyway everything seen so far ties up with the various presentations from Nvidia and IBM and one of those projects, and so far it looks like Pascal was a risk milestone/stepping stone specifically for Volta, hence why they did the unusual product launch cycle with Pascal but also paid off by providing high revenue.
But this is something I have said regarding the strategy/timeframe for some time now.
Cheers

Just for complete something.. Pascal architecture was unveiled in March 2015 ... And if you was remember it was on the exact same day they have present the TitanX ( Maxwell ). https://www.extremetech.com/gaming/...mance-gains-from-upcoming-pascal-architecture

 

[lanek]

Therefore, I, too, would be surprised to see consumer GV104 before 2018. Nvidia just doesn't need to. If pressured, Nvidia could easily release a fully enabled GP102 (maybe even with faster 11-12Gbps GDDR5X) in order to finish out 2017 on top (and they've shown that halo performance crown is important to them).

Well, the 1080TI spec will push nearly no difference on performance vs TitanX ( outside memory pool maybe 10 vs 12GB ), and this is with stock clock projected at <1600mhz (98-99%compute performance)... try imagine custom AIB GPU with 1700-1800mhz base clock. they will be way faster than TitanXP and for way less.

This is mostly the reason we dont have seen the 1080TI before, along that they had the 1080 ofc and AMD had only decided to release smaller gpu's.... In fact, in the price/lineup, between the 1080TI and TitanXP, one card is too much...

 

[CSI PC]

Therefore, I, too, would be surprised to see consumer GV104 before 2018. Nvidia just doesn't need to. If pressured, Nvidia could easily release a fully enabled GP102 (maybe even with faster 11-12Gbps GDDR5X) in order to finish out 2017 on top (and they've shown that halo performance crown is important to them).

If they operated like that you would not even had seen the Titan X in 2016, let alone in beginning of August as they did - Quadro and Tesla models were available around September-October time so not needed for those segments.
Also it is looking pretty certain consumer GP102 is going to be fully fleshed out very soon (already is on Quadro/Tesla): http://www.guru3d.com/news-story/nvidia-launches-geforce-gtx-1080-ti-countdown.html
There will be no more Pascal cards required in any of the ranges after that, every segment (retail/Tesla/Quadro) will be complete.
Cheers

 

[ImSpartacus]

It looks like Nvidia is going to rain on AMD's Vega parade with an early Volta debut in Q3 2017.

http://wccftech.com/nvidia-geforce-20-volta-graphics-card-q3-2017/

Wccftech is claiming the following GPUs will eventually make an appearance after GV104 in Q3.

• Volta GV100 GPU (Enterprise/HPC Only)
• Volta GV104 GPU (GTX 2070, GTX 2080)
• Volta GV102 GPU (GTX 2080 Ti / GTX TITAN XV?)
• Volta GV110 GPU

Part of me hopes that Nvidia has the balls to debut GV102 as "Titan Xv". The absolute madman.

 

[McHuj]

I hope so, I upgraded to a Gsync 144Hz monitor and now my 1070 seems so insufficient.

 

[gamervivek]

It'd be great if nvidia are returning the 600mm2 chip to desktop with GV110. Any hints as to whether Volta would be a repeat of Maxwell with lack of FP64 and what comes after it?

 

[CSI PC]

It looks like Nvidia is going to rain on AMD's Vega parade with an early Volta debut in Q3 2017.

http://wccftech.com/nvidia-geforce-20-volta-graphics-card-q3-2017/

Wccftech is claiming the following GPUs will eventually make an appearance after GV104 in Q3.



Part of me hopes that Nvidia has the balls to debut GV102 as "Titan Xv". The absolute madman.

I would expect the same run we saw with Pascal; meaning GV104 Q4 (still would think it is this but maybe at the beginning of the quarter if everything is going well) and a Titan that could be just shy of 3 months later.
May remember I mentioned in past Volta Xavier was meant to be according to some Q3-Q4 for sampling to manufacturers so maybe that did hit the Q3 schedule.
And it could be plausible this would also drag forward the other product lines if they wanted *shrug* ; we know Nvidia was pretty aggressive with the Pascal rollout across all diverse product families.
Cheers

 

[Jupiter]

SK Hynix: GDDR6 for new high-end graphics card early 2018: https://translate.google.com/transl...17-04/sk-hynix-gddr6-2018/&edit-text=&act=URL

 

[ImSpartacus]

SK Hynix: GDDR6 for new high-end graphics card early 2018: https://translate.google.com/translate?sl=de&tl=en&js=y&prev=_t&hl=de&ie=UTF-8&u=https://www.computerbase.de/2017-04/sk-hynix-gddr6-2018/&edit-text=&act=URL

Very interesting.

If the GV104-powered 2080 arrives in Q3 2017 as per this rumor and roughly matches the performance of the outgoing 1080 Ti (as historical G@104-400 parts have), then it needs roughly 450-500 GB/s of bandwidth.

That's attainable on a 256-bit bus (another G@104 classic) with 14-16 Gbps VRAM (be it GDDR5X or GDDR6). However, if 16 Gbps GDDR6 is coming in early 2018 and 14+ GDDR5X isn't coming soon (not on any recent roadmaps to my knowledge), then maybe a pair of HBM2 stacks are necessary?

Two stacks of 2 Gbps HBM2 would get the necessary 512 GB/s while also "cleanly" achieving 8GB of capacity (in 4-hi) in the 2080 with room to get 16GB (in 8-hi) for pro products.

After all, remember the old slides showing "stacked memory" was supposed to debut in Volta?

EDIT Just double checked Sammy and Micron. While I think Samsung's latest guidance is still a 2018 arrival for gddr6, I wasn't aware that Micron is getting aggressive and aiming to ship gddr6 in the 2017 calendar year. That could mean that GV104 could ride with a 256-bit bus one more time if a 2017 release is to be believed.

 

[Deleted member 13524]

What's the practical difference between GDDR5X and GDDR6?

 

[ImSpartacus]

What's the practical difference between GDDR5X and GDDR6?

I'd say the practical difference is that it's much easier for GDDR6 to get to 14-16+ Gbps whereas GDDR5X seems to have stalled around 10-12 Gbps due to it being an intermediary tech.

But I haven't found a good write-up explaining the more technical reasons for that practical result. I had the same question after Samsung announced their GDDR6 back at Hot Chips (and I don't think they've updated that 2018 roadmap), and no one around here seemed to have any answers.

https://videocardz.com/63391/hot-chips-26-first-details-about-hbm3-and-gddr6

 

[CSI PC]

Still more likely IMO that GV104 is GDDR5X and GV102 is GDDR6, while GV100 is HBM2.

There is no guarantee to the rating GDDR6 will initially launch at, look at the headaches of getting GDDR5X up to the expected rates.
Also worth noting Nvidia will be doing their memory contracts probably with Samsung and Micron rather than SK Hynix.

Cheers

 

[Deleted member 13524]

I'd say the practical difference is that it's much easier for GDDR6 to get to 14-16+ Gbps whereas GDDR5X seems to have stalled around 10-12 Gbps due to it being an intermediary tech.

But I haven't found a good write-up explaining the more technical reasons for that practical result.

Maybe GDDR6 is simply taking advantage of newer manufacturing nodes to get to higher clocks and use lower voltages here and there? I doubt they'll double the prefetch words again.

 

[Anarchist4000]

Maybe GDDR6 is simply taking advantage of newer manufacturing nodes to get to higher clocks and use lower voltages here and there? I doubt they'll double the prefetch words again.

That would be extremely likely and account for the performance and power savings. To really see gains would require an interesting change, and I don't believe GDDR6 is getting stacked yet like a lot of other memory techs.

 

[xpea]

GTC2017 will kick off in few hours with JHH opening Keynote but we already have some blurbs of what's coming.

By Inspur PR, we have definitive confirmation that Volta will be announced:

During GTC 2017, Inspur and NVIDIA and will jointly present a cutting-edge AI supercomputer, AGX-2, with the NVLink™ 2.0 enabled. The AGX-2 is designed to provide maximum throughput for superior application performance for science and engineering computing - taking AI computing to the next level.

Note: Pascal is only NVLink 1 and Volta is NVLink 2
source: http://www.prnewswire.com/news-rele...uter-and-aistation-at-gtc-2017-300452957.html

SK Hynix is diplaying GDDR6 in their booth:

More info at the source: http://www.tweaktown.com/news/57460/sk-hynix-teases-gddr6-gtc-2017-16gbps-bandwidth/index.html

and finally the bomb dropped by JHH him self during Q1 record earning QA session:

C.J. Muse - Evercore Group LLC
Very helpful. I guess as my follow-up, on the inventory side, that grew I think 3% sequentially. Can you walk through the moving parts there? What's driving that, and is foundry diversification part of that? Thank you.

Jen-Hsun Huang - NVIDIA Corp.
The driving reasons for inventory growth is new products, and that's probably all I ought to say for now. I would come to GTC. Come to the keynote tomorrow. I think it will be fun.

Nvidia is already stock pilling Volta...

 

[Pressure]

Wow, they just "uneiled" the first GDDR6 How did that slip through the proof reading.

 

[DavidGraham]

How are they stockpiling Volta right now? I thought it's gonna be 12nm, don't think that process is mature enough yet, right?

 

[Deleted member 13524]

Nvidia is already stock pilling Volta...

GP108 cards (Pascal GT 1030) haven't been announced yet. It could also be an Android console or set-top box using Parker. Or even just the FCC-approved Shield handheld with a TX1).

Volta is using TSMC's 12nm. Its launch date, according to SK Hynix themselves, is early 2018 with a 384bit GDDR6 arrangement (at least the consumer GV102 variant).
I have no doubt production for Volta will start in 2017, but I find it hard to believe they've been building stock for it during Q1 2017 (which is what the question refers to).

 

[xpea]

of course, they are stock pilling HPC Volta GV100 with HBM2, not the consumer version.