NVIDIA GeForce 20 Series (Volta) to be released later this year - GV100 announced

[xpea]

Why would they make a 1030 but no 1040? (Does 1040 sound too much like a tax form?)

Reason is simple: 4 is a bad number in China (some buildings don't even have 4th nor 14th floor). At the opposite, 8 is a lucky number

 

[Sweepr]

GeForce GT 1030 (leaked by MyDrivers.com) now confirmed and pictured by other chinese websites:

http://www.ithome.com/html/digi/305139.htm
http://www.expreview.com/53583.html

 

[Samwell]

Gt1030 is coming now and they might have some sources, but their Volta news just make no sense. You can't just push a product forward 6 month, if it's not already finished and then you would just plan it earlier. There are GV100 signs in drivers, but of no other Volta chips, so GV100 should come first and it's expected at the end of this year. Consumer volta won't come in this year.

 

[tamz_msc]

GeForce GT 1030 (leaked by MyDrivers.com) now confirmed and pictured by other chinese websites:

http://www.ithome.com/html/digi/305139.htm
http://www.expreview.com/53583.html

I thought that this thread was about Volta.

Anyway, it looks like the GT 1030 could be based either on the Quadro P400 or P600. If it is based on the P400(which I hope it isn't) then the RX 550 is twice as powerful in terms of theoretical TFLOPS. Less likely as the GK208 had more cores, but you never know, NVIDIA's low end is a confusing mess of cards of different generations labelled with the same name. If it is based on the P600, then it is on par with the RX 550in TFLOPS.

 

[OatisCampbell]

Agreed.



NVIDIA makes its money from high end gaming/data center/workstation products, chips like GM107 just aren't that important these days. Volta for consumer, if it comes this year, will be GV104, not GV107 or the like. Ask yourself, what came first, GP104 or GP107?

Yep. Next gen mid-range that outperforms 1080 Ti for $700..
People forget these chips are in development for years and there is no "Gosh competitor X is doing something, give me a new part with 50% more performance-NOW!"

This is why Vega is not out, this is why Volta may be out in some form this year.

All depends on if the part can be made profitably on current tech silicon. (and sometimes on where memory or other related tech is at)

A year and a half after Pascal would be on the fast side of normal release for NV, but if parts are almost ready their margins are fat and the only downside I can see is it would cannibalize that new Titan's sales. Which probably isn't meant to be a volume part anyway, and they'll have sold a ton of Tis by then anyway.

 

[SteveGrabowski]

Damn if we get Volta in 2017 I might upgrade my 970. That would be nice to see a 2060 with 1070 level performance for $250 or so.

 

[Cookie Monster]

I always thought of Volta as a new architecture from nVIDIA, something along the lines of a G80, Fermi jump (architecturally speaking of course) but perhaps it could be an enhanced Pascal. And thats not necessarily a bad thing.

Since GDDR5 and 5X consumes roughly 50~75W (say for a 384bit bus/memory controller) accordingly to my knowledge and given the 250W power envelope that current high end GPUs are trying to maintain.. a jump to HBM2 memory to make space within the given power envelope for more functional units like ALUs etc (perhaps along side with 12nm process) could essentially be what Volta is. Of course there will be minor updates within the architecture itself but the crux of the power savings could be the jump to HBM2 or stacked memory etc for their Geforce line.

And with the debut of a GV104 with HBM2 for instance, most of these products could be sitting at the top of the pecking order i.e. more $$$ or at the price the GTX1080 launched (it could be roughly x2 performance and similar ~180W TDP).

Just speculating of course.

 

[Ajay]

GDDR6, it seems to me to, renders HMB HBM pointless for consumer grade graphic cards.

 

[Cookie Monster]

GDDR6, it seems to me to, renders HMB pointless for consumer grade graphic cards.

Your probably right especially on the $$$ part given the performance pros and cons between the two.

 

[Samwell]

GDDR6, it seems to me to, renders HMB pointless for consumer grade graphic cards.

HBM2 it seems yes, but GDDR6 is at it's limits already with 16 Gbps. The generation after that will either need GDDR7, which probably won't release so fast or will use HBM3 which i believe. Then the technique should be old enough to produce it cheap.

 

[crisium]

I'm not familiar with the technical specifications. How do you know 16 Gbps is the limit? GDDR5 can do 9 Gbps today, should the successor not be able to double? Would a GDDR6X also be possible? We already have 11 Gbps of that.

Edit: Another thought is that we won't see the full 16 Gbs right away, by Nvidia choice. They could do lower, saving faster for a Volta refresh or Volta's successor. Nvidia tends to use the fastest available for their mid and high end cards (AMD tends to use cheaper, slower memory), but it is a possibility.

 

[DarthKyrie]

HBM2 it seems yes, but GDDR6 is at it's limits already with 16 Gbps. The generation after that will either need GDDR7, which probably won't release so fast or will use HBM3 which i believe. Then the technique should be old enough to produce it cheap.

Seeing as Intel has embraced HBM2 for their new G line with AMD GPU, HBM2 should come down in price.

 

[Harmaaviini]

Considering that GDDR6 is scheduled for late 2017/early 2018 if there is a Volta card coming in Q3/17 I really doubt it can match 1080 ti. 3 reasons why I think that: I doubt that nV is willing to increase the die size of GV104 that much. GDDR5x not having enough bandwith to do that on a 256bit bus anyway (right?). Yes, I'm assuming nV is going to maintain that width on their 1080 successor. Jump from 16nm finfet to 12nm finfet is not going to be nearly as big as it was from 28nm bulk to 16nm finfet.

About the possibility of Q3 launch for a Volta product: It would have had to been taped out late last year to make it in Q3 at all, right? Were there any rumors to indicate that at the time?

 

[Qwertilot]

It is deifnitely relatively plausible - both very big compute volta and the Tegra variant (for car stuff) are due on that kind of time scale. Not such a huge jump to GV102/4 from those two.

 

[Genx87]

I'd think Nvidia wants to pee in AMDs Cheerios with this launch. So Q3 seems weird to me. Is it possible they are confusing a Volta launch for a super computer with general consumer availability?

 

[caswow]

AMD has engineers working on gddr6 too you guys are fooling yourselfs believing amd will go gddr5 or hbm2 only.

 

[nathanddrews]

I don't care what RAM they use as long as it's fast enough to perform its function.

EDIT: I would like to see something like AMD's SSG come to consumer GPUs, where we get an M.2 slot on the GPU for massive, high-speed asset storage.

 

[Sweepr]

Fudzilla: Nvidia to showcase Volta in early May

Volta, the successor to the Pascal architecture, will first debut as a data center, deep learning, artificial intelligence product, since margins in this market are much higher than with the gaming stuff.

...
One logical codename for the GP100 codename would be the GV100, as in Geforce Volta 100. A few magazines pointed out that Volta might come in 2018 with GDDR6, but Fudzilla expects that we will see Volta based GPUs probably before the end of summer, or should we say Q3 2017.

...We expect to see Volta using HBM 2 memory for the GV100 part and the GV104 - the GPU version of the chip is likely to use the GDDR5X with some Titan versions possible with HBM 2, but only if being really pushed by Vega.

Of course, Jensen will shortly share some official Volta information, in the second week of May.

www.fudzilla.com/news/graphics/43485-nvidia-to-showcase-volta-in-early-may

 

[antihelten]

Fudzilla: Nvidia to showcase Volta in early May

www.fudzilla.com/news/graphics/43485-nvidia-to-showcase-volta-in-early-may

So no GDDR6 for either GV100 or GV104 then? I wonder if the forthcoming 384bit GPU that SKHynix was talking about is the GV102 card, which could then arrive in 2018?

So the lineup could then be:

GV100 - HBM2, 6144 bit (6 modules), 1075-1536 GB/s, 24GB VRAM - Q3 2017
GV104 - GDDR5X, 256-384 bit, 352-576 GB/s, 12GB VRAM - H2 2017
GV102 - GDDR6 - 384 bit, 672-768 GB/s, 12GB VRAM - H1 2018

On the other hand this line makes it rather apparent that Fudzilla is probably just pulling this from where the sun don't shine:

the GPU version of the chip is likely to use the GDDR5X with some Titan versions possible with HBM 2

Seeing as you obviously can't just make a different version of the chip with HBM2. That would have to be a completely new design and thus a completely different chip.

 

[crisium]

Two things wrong with Fudzilla here. One is GV104 in HBM2 and GDDRX. Nonsense, as you point out.

The other thing is GV104 with Titan versions. That's what GV102 will be for.

But wait.

What if, 4 flagships were not enough? What if they need more? Envision if you will:

GTX 2080 flagship (cut GV104) followed by
Titan XV flagship (full GV104) followed by
Titan Xv flagship (cut GV102) followed by
GTX 2080 Ti (cut GV102 but faster) followed by
Titan XXX flagship (full GV102)

Don't say they wouldn't do it. =P

 

[Arachnotronic]

In this discussion, you should all keep in mind that NVIDIA will likely release a 7nm GPU architecture in early 2019.

 

[antihelten]

In this discussion, you should all keep in mind that NVIDIA will likely release a 7nm GPU architecture in early 2019.

That would still leave plenty of room for a full Volta lineup on 12/10 nm before then, assuming that Nvidia starts launching it in Q3 2017. In fact that would be close to the same time frame that Pascal is getting.

 

[Arachnotronic]

Two things wrong with Fudzilla here. One is GV104 in HBM2 and GDDRX. Nonsense, as you point out.

The other thing is GV104 with Titan versions. That's what GV102 will be for.

But wait.

What if, 4 flagships were not enough? What if they need more? Envision if you will:

GTX 2080 flagship (cut GV104) followed by
Titan XV flagship (full GV104) followed by
Titan Xv flagship (cut GV102) followed by
GTX 2080 Ti (cut GV102 but faster) followed by
Titan XXX flagship (full GV102)

Don't say they wouldn't do it. =P

Doubt it.

I think we get full GV104 in Q3ish, full GV102 in the form of a Titan Xv in Q1, and then a GV102-lite in Q2.

 

[Bouowmx]

For 7 nm, Volta shrinked? Like Maxwell shrinked for 16 nm.

 

[Arachnotronic]

For 7 nm, Volta shrinked? Like Maxwell shrinked for 16 nm.

No idea. I'm sure they'll call it something (Newton?) and make some architectural changes like they did with Pascal, but I think Volta is the "tock" to the (first) 7nm product's "tick."