NVIDIA Corporation Volta Architecture Rumor Emerges

razor1 said:
nV needs to have cards that their users upgrade to too.

nV's competition currently is themselves.
Click to expand...

Well said. For me the availability of high refresh rate 4K displays is also a big factor in upgrading. I am perfectly happy with the performance I am getting from my 980 Ti for 1440p @ 144 Hz and since the displays I mentioned seem to come out earliest late this year, I will probably wait until next year to upgrade. On one hand I really wish I could hop on that train sooner rather than later but on the other hand I am pretty happy how much value I've got from my 980 Ti investment.
 
euskalzabe said:
Um, what? Are people smoking something new these days? I've seen this kind of misinformation in a couple other threads today. To be clear: Volta was supposed to be released in 2016. Then big Maxwell took a few months too long, and Nvidia realized stacked RAM was way too hard for 2016/2017. So Pascal came up - essentially a Maxwell node shrink + adaptations. What they call a "new architecture" is what Ivy Bridge was to Sandy Brigde, what Broadwell was to Haswell - a tick in Intel nomenclature, aka node shrink. Nothing more. If there's a Pascal "refresh" - cos AMD ain't competing much right now - it's more of a Pascal rehash - nothing new, sped up clocks, call it a day. We just did the shrink with Pascal, next should be actual new architecture. So, Volta should come out in 2017 with stacked RAM as was promised years ago - anything else released, tells you something went wrong, but Nvidia isn't one to admit and own up to its mistakes. Either way, depending on how you look at it, it's at least 1 year late:
Click to expand...

Not sure why people use product roadmaps as an explicit launch cycle (people did the same with AMD and Vega), not directed at you but have this conversation with others on some sites.
Volta was always 2017 even back in 2014.
The 2014 whitepaper-technical paper for the 1st Tesla Volta dGPU accelerator mentions 2017, every HPC Tesla presentation I have seen also mentions 2017.
The end of 2014 project update for the CORAL implementations has the timeframe as 2017 for the acquiring of Volta, these are the 1st supercomputers to use Volta.
It seems to me Pascal was brought in primarily as a technical/risk milestone to ensure not everything was brought in one go for the V100 as that would be way too many new technical features (big ones introduced with Pascal being HBM2 for ther 1st attempt, node shrink,unified memory,NVLink,mixed precision,improved engineering with the node properties-power envelope and propertied,) ; one reason to think this is the very 1st dGPU was the P100 and this would be very expensive with risks compared to the normal approach of starting small and end with the largest die.
Yeah of course it also means revenue now without waiting a whole year and is a large consideration, but then that would not need to start with the highest risk approach of a 610mm2 die and the P100 with the advanced features it has compared to consumer dGPU.

The biggest risk to Nvidia business growth-strategy for now is Intel, and a few others when branching outside HPC-data analytics into newer opportunities.
Cheers
 
Last edited:
Take with a grain of salt and skepsis as usual.

GV104 from chinese forums delivers 13.5tflops. As compare GP104 delivers 8.9tflops and GP102 in the Titan X 11tflops. (Theoretical tflops for all).

That's a major uarch rework like Maxwell if true.
 
Shintai said:
Take with a grain of salt and skepsis as usual.

GV104 from chinese forums delivers 13.5tflops. As compare GP104 delivers 8.9tflops and GP102 in the Titan X 11tflops. (Theoretical tflops for all).

That's a major uarch rework like Maxwell if true.
Click to expand...

That's a 50% increase. Since density remains largely the same from the looks of it, even higher clocks? GV104 being a larger chip than GP104 seems unlikely, since that'll drive down margins.
 
trandoanhung1991 said:
That's a 50% increase. Since density remains largely the same from the looks of it, even higher clocks? GV104 being a larger chip than GP104 seems unlikely, since that'll drive down margins.
Click to expand...

GK104(3.1Tflops) was 294mm2 and GM204(4.6Tflops) was 398mm2. Both using 28nm. GV106 will be made with a 4th generation 16nm node, now called 12nm. GP104 was made with a second generation 16nm node.

There is obvious more than that to it.
GK104 1536:128:32 config and 195W
GM204 2048:128:64 config and 165W

Margins wont be lower just for that. There is a lot more to it than that. Else by now AMD and NVidia should sell 100mm2 top bins? :p

I forgot to add, the rumour also says GV104 will be paired with 16GB VRAM. (12-14Ghz GDDR5X/GDDR6 I guess).
 
Last edited:
They may be speculating-basing the figures of what has been suggested by Nvidia/IBM relating to Volta V100 and how that compares to the P100 on a per node basis.
But it makes sense the 1st consumer GPU would be GV104 as Nvidia will be launching same approach as Pascal.
I still think Nvidia would be deciding whether to release GV104 late this year (November-Dec launch absolute earlierst IMO) or next year as there are benefits to each approach, it would kill Vega in terms of margins from both manufacturing and rrp with its relative performance and put AMD under serious pressure in the above mainstream market.
But it would also wipe out sales of a 1080ti, still that is a niche product compared to how well a GV104 will sell as many would upgrade with that kind of performance again and the gap between them would still be 6ish months (if we see a 1080ti) worth of 1080ti revenue.
Also of interest would be the GV104 will probably follow same trend of FE models, and that did earn Nvidia a fair chunk of money.

Cheers
 
Last edited:
I don't see it happening this year. GDDR 6 is suppose to hit next year, unless Nvidia goes to a higher bandwidth bus for GV 104 which I doubt - it will be next year. I do see the 1080Ti coming out and some modest updates or refresh to Pascal is possible. A higher end Titan could also come into play. 1080Ti, $799, Current Titan X $999 and maybe fully enabled Titan Black? $1200
 
CSI_PC said:
They may be speculating-basing the figures of what has been suggested by Nvidia/IBM relating to Volta V100 and how that compares to the P100 on a per node basis.
But it makes sense the 1st consumer GPU would be GV104 as Nvidia will be launching same approach as Pascal.
I still think Nvidia would be deciding whether to release GV104 late this year (November-Dec launch absolute earlierst IMO) or next year as there are benefits to each approach, it would kill Vega in terms of margins from both manufacturing and rrp with its relative performance and put AMD under serious pressure in the above mainstream market.
But it would also wipe out sales of a 1080ti, still that is a niche product compared to how well a GV104 will sell as many would upgrade with that kind of performance again and the gap between them would still be 6ish months (if we see a 1080ti) worth of 1080ti revenue.
Also of interest would be the GV104 will probably follow same trend of FE models, and that did earn Nvidia a fair chunk of money.

Cheers
Click to expand...

Yep, GV104 comes first for gamers. That's 100% certain. But GV100 will be the first one to launch.

noko said:
I don't see it happening this year. GDDR 6 is suppose to hit next year, unless Nvidia goes to a higher bandwidth bus for GV 104 which I doubt - it will be next year. I do see the 1080Ti coming out and some modest updates or refresh to Pascal is possible. A higher end Titan could also come into play. 1080Ti, $799, Current Titan X $999 and maybe fully enabled Titan Black? $1200
Click to expand...

10Ghz isn't the limit for GDDR5X if that's what you think about. On paper its 14Ghz for GDDR5X and 16Ghz for GDDR6. Also remember how GDDR5X wasn't supposed to exist at the time. And HBM2 as well it seems.
 
Last edited:
noko said:
Well if Nvidia has another pleasant surprise that will be interesting - Raja may hang himself afterwards.
Click to expand...

I dont think he cares a single bit. And specially not as long as he is paid.
 
noko said:
I don't see it happening this year. GDDR 6 is suppose to hit next year, unless Nvidia goes to a higher bandwidth bus for GV 104 which I doubt - it will be next year. I do see the 1080Ti coming out and some modest updates or refresh to Pascal is possible. A higher end Titan could also come into play. 1080Ti, $799, Current Titan X $999 and maybe fully enabled Titan Black? $1200
Click to expand...
The 12Gb/s GDDR5x has gone into production status now, so that is another 20% increase on what has been used so far and there is still 14Gb/s in the DRAM product part papers of Micron (albeit not even in sampling status yet), plus there is no real guarantee when GDDR6 will go into sampling status.
I agree that would be the ideal memory for mainstream-performance but depends if there can be any guarantees for sampling and production status timeline, and how that fits into product launch cycles.
Cheers
 
Last edited:
Shintai said:
Yep, GV104 comes first for gamers. That's 100% certain. But GV100 will be the first one to launch.
Click to expand...
Yeah,
Intel is their motivator and also the contract obligations Nvidia are signing for the various supercomputer projects, the HPC/massive nodes data-modelling/Analytics/Deep Learning/auto'transport' is creating a new tech arms race not just with these 2 companies but also more specialised ones trying to get a piece of the action.
Nothing against AMD as they have some nice ideas-tech coming along but they have a massive task to compete in even just a few of these markets and some pieces are still missing for them.
Cheers
 
Last edited:
So the question becomes how ready is GV 104? We don't know other then rumors. Agree GDDR 5X could be use and probably will be fast enough as well using the higher speed 12GB/s+ now available for the GTX 2080 and maybe slower 10Gb/s for the 2070.

The 2080 would most likely be as fast or faster then the 1080Ti which would cause a conflict if released. If the 1080Ti is not released then NVidia has an inventory of GP 102 GPU's that could be used for the 1080Ti. I really don't know why Nvidia is holding off on the 1080Ti at this point if Volta consumer will be coming out this year - now would be the time to get it out.

I will still be very surprised if Nvidia will launch Volta this year for the consumer - unless it is very much ready - Vega really beats Pascal significantly and if it does not conflict with the HPC or GV 100 production. Nvidia could just kill AMD with pricing, move out the inventory and then launch Volta at a state AMD could not effectively compete next year. I would not get my hopes up too much for Volta this year for the consumer.
 
CSI_PC said:
Yeah,
Intel is their motivator and also the contract obligations Nvidia are signing for the various supercomputer projects, the HPC/massive nodes data-modelling/Analytics/Deep Learning/auto'transport' is creating a new tech arms race not just with these 2 companies but also more specialised ones trying to get a piece of the action.
Nothing against AMD as they have some nice ideas-tech coming along but they have a massive task to compete in even just a few of these markets and some pieces are still missing for them.
Cheers
Click to expand...
I would also think that GV 100 would effectively stop Vega in the HPC market as well. AMD is working on some rather dense Volta options putting more compute power in a given volume so AMD is using their options as best they can to get involved more with the HPC market. Here is Kyle's buddy holding up a plastic cube with 4 Vega GPU's :smug:

forum image hosting
 
GV100 should be announced at GTC in april tho actual products will be Q3 or so. 1080TI in march according to rumours. So if we are really lucky we can get a GV104 out in Q4.
 
Shintai said:
GV100 should be announced at GTC in april tho actual products will be Q3 or so. 1080TI in march according to rumours. So if we are really lucky we can get a GV104 out in Q4.
Click to expand...
The 1080Ti sounds right but new line of cards in Q4 does not. Good time to dump inventory though. GV 100 maybe Q4 reason being the extreme amount of testing that is needed before committing to some rather expensive systems not only in hardware but everything else. If GV 100 was out with engineering samples right now with respective interested groups to test for HPC systems, software etc. I would think Q4 would be possible.
 
noko said:
The 1080Ti sounds right but new line of cards in Q4 does not. Good time to dump inventory though. GV 100 maybe Q4 reason being the extreme amount of testing that is needed before committing to some rather expensive systems not only in hardware but everything else. If GV 100 was out with engineering samples right now with respective interested groups to test for HPC systems, software etc. I would think Q4 would be possible.
Click to expand...
It would not necessarily be a new line of consumer cards, they could just launch a 1080 replacement, same time or 1 month later a 1070 replacement, 4-5 months later a lower card and others following on from that.
The Maxwell 960 came out 4 months after the 980/970.
But yeah I would not want to bet either way with regards to when we will see the 1080 replacement and whether they launch the cut die GV104 at the same time; They may allow a mix of Pascal (priced adjusted and allow control clearance of stock) and upper Volta at the same time
Plenty of options for Nvidia from a business-timing strategy perspective.
Just feels logical they would not want to wait 6-8 months for launching a consumer Volta after a Tesla Volta as it really messes product cycle roadmap and development-business strategies and also Nov-Dec is a perfect time for a 1080 replacement beyond reasons already given in earlier posts as it also aligns with the Christmas season.
All depends if Nvidia decide to use the '12nm' that is not 12nm node for the GV100 and how that is also for production time - should not be much of a risk as it is in reality another option as part of the 16nm family offered by TSMC and does not even have its own R&D roadmap.
Cheers
 
Last edited:
Guess Guess guess, sounds like AMD threads about Vega. As for Vega, in May, the usual dog and pony show, June first models available (they run out), Aug availability is not bad, new models etc. Pascal will be doing great rest of year. Volta GV 104 could be basically a bolt on GPU using faster available memory, partial refresh partial next generation. HBM higher end versions come later. I just don't see it happening is all this year. Hopefully I am wrong.
 
noko said:
Guess Guess guess, sounds like AMD threads about Vega. As for Vega, in May, the usual dog and pony show, June first models available (they run out), Aug availability is not bad, new models etc. Pascal will be doing great rest of year. Volta GV 104 could be basically a bolt on GPU using faster available memory, partial refresh partial next generation. HBM higher end versions come later. I just don't see it happening is all this year. Hopefully I am wrong.
Click to expand...
I would not underestimate Volta, it is looking close to another 50% speed gain going by what has been mentioned indirectly by Nvidia/IBM regarding Tesla Volta compared to Tesla Pascal.
Cheers
 
CSI_PC said:
I would not underestimate Volta, it is looking close to another 50% speed gain going by what has been mentioned indirectly by Nvidia/IBM regarding Tesla Volta compared to Tesla Pascal.
Cheers
Click to expand...


Yeah that's for unit for unit, so depending on clock speed that could be different too, if they can get more clocks out of the new architecture it might be more than that when we look at gen to gen of equivalent segment cards.
 
Well we all know where the money is. I think Vega probably is a big step forward in the server market when it comes to performance/efficiency. Good for them, if they can make up some ground there it will translate in to more money for R&D. I have no problem with Vega as long as its decent in gaming. I don't know if I will be buying one, thats up in the air. Really depends on the price/performance. My monitor does have freesync so I might be temped. But like I said I am ready to be underwhelmed like always but I am sure it is decent. Ryzen will be what makes or breaks AMD though. GPU division has been decently competitive up until last year, yea not the best but still decent despite the lack of funding.
 
I was waiting on Vega and then I found an MSI RX 480 for $155 ($175 - $20 rebate), so I though why the heck not... even if it only lasts a year, that's great value. I'll be intensely following the Vega release though, all this memory talk has me much more intrigued than Volta's efficiency.
 
trandoanhung1991 said:
That's a 50% increase. Since density remains largely the same from the looks of it, even higher clocks? GV104 being a larger chip than GP104 seems unlikely, since that'll drive down margins.
Click to expand...

Margins assumes the SKU retails for the same price as current SKUs which it may totally not depending on the competitive landscape.

We've seen steady creeping up in prices for "mid-range" x04 chips, why not expect the same trend to follow through?

GV104 "Founder's Edition" for $899 anyone?

Margins will be even better, as this "12nm" is just an improved 16ff which by then is mature and high yielding.

2017 though, I expect a Pascal refresh, akin to the 680 -> 770 situation. GTX 2080 refresh, GP104, higher out of the box boost clocks, at a slightly lower price, say, a more sane $499 would sell very well.
 
Anyone think Nvidia will hold back launching Volta until next year need to consider they are being pushed by Intel rather than AMD.
What drives Nvidia in the HPC space will move it quicker as well in the consumer segment as well, it is pointless for them to launch the design for HPC and wait 6-9months to then launch for consumer market as they would end out of synch launch cycle wise and the GPU designs are needed for both.
I mentioned there is a new tech war race in the HPC/Deep-learning space and this will dictate launch cycles.
Beyond Xeon Phi and associated Knights Landing with Altera, seems Intel is moving at a pretty good pace with their Nervana acquisition (this will push Nvidia).
Intel Nervana Artificial Intelligence Meetup 1/31/17:www.slideshare.net/nervanasys/intel-nervana-artificial-intelligence-meetup-13117

Edit:
Worth noting Nervana is considered for both its own hardware and also next gen Xeon Phi.
Cheers
 
Last edited:
horrorshow said:
If Volta launches before 2018, I'll grow a third arm.
Click to expand...
Then they start to lose out to Intel in terms of HPC tech momentum, bear in mind this is what drove the launch of Pascal not consumer; consider what launched 1st an HPC card using the largest die possible that adds greatest cost and risk compounded by some 1st time tech for Nvidia (not just HBM but NVLink/unified memory-page faulting/offload mode/etc).
Cheers
 
It's just common sense is all I'm saying.

How often do you see a new GPU architecture less than 18 months apart?

ESPECIALLY if you're planning to shrink the die.

New fabrication process etc etc.

I'm all for speculation but I think some of you guys are getting excited a bit too early....

EDIT: If I were a betting man, I'd wager we'll see the 1080ti in April/May.
 
horrorshow said:
It's just common sense is all I'm saying.

How often do you see a new GPU architecture less than 18 months apart?

ESPECIALLY if you're planning to shrink the die.

New fabrication process etc etc.

I'm all for speculation but I think some of you guys are getting excited a bit too early....

EDIT: If I were a betting man, I'd wager we'll see the 1080ti in April/May.
Click to expand...


nV's launch cycle has been 18 months since the g80 and they only missed one time (Fermi) by more than one quarter since.....
 
razor1 said:
nV's launch cycle has been 18 months since the g80 and they only missed one time (Fermi) by more than one quarter since.....
Click to expand...

Yeah, that's correct.

But between HBM2 and the die shrink, I seriously doubt they'll make Q4.
 
horrorshow said:
It's just common sense is all I'm saying.

How often do you see a new GPU architecture less than 18 months apart?

ESPECIALLY if you're planning to shrink the die.

New fabrication process etc etc.

I'm all for speculation but I think some of you guys are getting excited a bit too early....

EDIT: If I were a betting man, I'd wager we'll see the 1080ti in April/May.
Click to expand...
The '12nm die' shrink is not really a node shrink as such, it is an option to the 16nm family from TSMC, hence it is a low risk option and one reason it has never shown up on TSMC R&D roadmap.
Volta was always planned as 16nm.

Nvidia is already using HBM2 and has been since summer last year.
Pascal is a tech-risk milestone for the HPC 'V100' and also Volta, look at what launched 1st and with the 1st iteration of tech that is going to evolve for Volta.
Yes it also provided nice revenue but it is pretty clear from their launch strategy what their focus is, look at the focus Nvidia did for Pascal with Drive PX2 and P100, exactly same period a year later we get the update on Xavier to replace Drive PX2 and expectation of the V100 details at GTC in May this year (P100 was introduced/launched GTC 2016).
They are being pushed by Intel rather than AMD, and a few of the other more specialist companies in their specific fields of expertise/segments.
Cheers
 
Last edited:
horrorshow said:
Yeah, that's correct.

But between HBM2 and the die shrink, I seriously doubt they'll make Q4.
Click to expand...


They are already on HBM2 on Pascal, so that's not an issue

And 12 nm is 16nm, it doesn't even look like a half node change, just tweaks of 16nm from the look of it.
 
Not on consumer cards.

Maybe, just maybe we can expect the new Titan Q4.

Again, maybe....
 
putting HBM 2 on a consumer GPU is pretty much the same thing as another other memory tech, just needs the HBM2 memory bus and interposer. Now the reason nV hasn't done it for their consumer cards is they don't need to yet, GDDR 5x has been enough (which they haven't used the fastest versions of GDDR 5x yet) So depends of they need a lot more bandwidth than GDDR5x can provide.
 
Only reason I posted in the first place is your timetable is very optimistic.

If I'm wrong, stand on me.

But, I'm fairly sure I'm right...
 
Volta was always late 2017 early 2018 product, even on their time lines. 10nm was pushed out, but 16nm still has space for them to do what they need to do for consumer GPU's. Tesla products with full DP different story though.
 
You must log in or register to reply here.
Back
Top