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

[raghu78]

AMD is planning to move to a new generation (Polaris -> Vega) in about 13 months. If AMD can do it, I don't see why NVIDIA can't pull it off, too.

Also, there was a lot of derision of Pascal as "Paxwell" because the underlying uArch was very similar to Maxwell's (with some enhancements & implemented in 16FF+), so it's not too farfetched to expect NVIDIA to move to a "true" new architecture 15 months-ish after Pascal.

Frankly imo Vega with NCU is the first real update to the GCN architecture in 5 years. Almost every subsystem including the command processor, the compute unit, the geometry processor and the rasterizer has been improved. the Up until now all AMD was doing was minor tweaks to GCN like color compression (introduced in Tonga and improved in Polaris), improved tesselation (first introduced in Tonga and improved in Polaris), Primitive discard accelerator(introduced in Polaris). The core engine has not changed much and due to that reason we could see Fiji hitting serious bottlenecks.

As for Pascal it had some nice improvements to Maxwell like async compute but the base architecture was already very very good. Still Nvidia did a fantastic job with improving clocks and with the TSMC 16FF+ process hit 2 Ghz clocks easily (the first ever for a GPU). imo Volta is probably the first major architecture change since Fermi (which was built for DX11). I think its going to be built grounds up for modern APIs like DX12 and Vulkan. I think professional Volta with IBM Power 9 will launch for supercomputer customers in late 2017 and consumer Volta will come in late Q1 or early Q2 2018. If we do see Geforce Volta in 2017 it will probably be a low power chip (<75w) like GM107 for notebooks and entry level gaming.

 

[railven]

My comment was in relation to what the Chinese article said about the Chinese market. If they are seeing less demand even with a GTX 1080 price cut and surprisingly competitive pricing of the GTX 1080, then I don't know how they'll increase demand by releasing new products at higher prices.

My guess would be pushing the prior product stack down in regards to price. They satisfy demand on the bottom while creating cash flow on the top. As I alluded to in my other comment "I want it now price be damned" versus "I can wait for a discount" buyers. With the ultimate goal of making money, NV has no reason to reduce prices. I wouldn't consider China's demand to be for the top.

 

[Arachnotronic]

As for Pascal it had some nice improvements to Maxwell like async compute but the base architecture was already very very good. Still Nvidia did a fantastic job with improving clocks and with the TSMC 16FF+ process hit 2 Ghz clocks easily (the first ever for a GPU).

Agreed.

I think professional Volta with IBM Power 9 will launch for supercomputer customers in late 2017 and consumer Volta will come in late Q1 or early Q2 2018. If we do see Geforce Volta in 2017 it will probably be a low power chip (<75w) like GM107 for notebooks and entry level gaming.

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?

 

[raghu78]

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?

Actually GM107 is the more relevant example as it involved a move to a significantly improved architecture from Kepler to Maxwell on a mature 28nm process. PaScal was less of an architectural change compared to Kepler->Maxwell even though it involved a move to a brand new FINFET process ,16FF+. Anyway Nvidia has shown they can take different approaches when moving to a new generation. So what path they choose remains to be seen.

 

[DooKey]

It wouldn't surprise me one bit if Volta makes an appearance in time for the Christmas holiday buying period. I hope it does come then, but I'm not going to hold my breath.

If it does come out then I'll move my Titan X(16) to the Intel box in my sig (wife's computer) and buy a GV 104 for my system since it will probably give 20% uplift over the Titan X if last generation is any kind of indicator.

 

[Ajay]

So, more info on Volta being designed to use TSMC's 12FF node (improved 16FF+) from Digitimes: http://www.digitimes.com/news/a20170313PB201.html

TSMC lands orders for HPC chips from Nvidia, Qualcomm, says paper
Commercial Times, March 13; Steve Shen, DIGITIMES [Monday 13 March 2017]
Taiwan Semiconductor Manufacturing Company (TSMC) has landed orders for high performance computing (HPC) chips for AI applications from Nvidia and Qualcomm, respectively, according to a Chinese-language Commercial Times report.

TSMC will fabricate Nvidia's next generation Volta GPU using a 12nm process, said the paper, which added that the Volta GPU will be paired with Nvidia's Xavier supercomputer chips for self-driving car applications.

 

[Sweepr]

Usually I would have scoffed at this, but Micron fast tracking GDDR6 to this year says it all. Not sure how I feel about this, but progress is progress.

This is bad news for AMD, as it looks like they won't get any traction this year from Vega, unless Vega turns out to have higher performance than the 1080 Ti; which is unlikely in my opinion.

I'm not sure how they're going to respond if NVIDIA manages to deliver >20% better performance than GP102 at <200W TDP, assuming the first card is GTX 1080's successor (GV104?). Vega 10 needs to improve 1440P/4K gaming performance by >70% compared to Fury X to barely match 1080 Ti.

As for the people questioning the source. MyDrivers broke the news about low-end Geforce GTX 1030 a week ago. Now PConline confirmed the card does exist. The same website leaked info about AMD's plans to refresh their mainstream cards with revamped Polaris GPUs (also confirming Polaris 12 at the occasion) and even got the launch date right (April 18) more than a month ago - all the usual websites quoted them (1, 2).

 

[tviceman]

I'm not sure how they're going to respond if NVIDIA manages to deliver >20% better performance than GP102 at <200W TDP, assuming the first card is GTX 1080's successor (GV104?). Vega 10 needs to improve 1440P/4K gaming performance by >70% compared to Fury X to barely match 1080 Ti.

Assuming we're looking at 16nm/12nm I'm guessing Volta GV104 will be to GP102 that Maxwell GM204 was to GK110 - about 10% faster, which will still make it impossible for AMD to compete with Vega being way bigger in die size, way more power hungry, and strapped with way more expensive HBM2 memory. Poor AMD; they should have reverted back to HD 300/400/500 series cards and stuck with making the best possible mid-range dies possible. The 4850/4870 and 5850/5870 were very robust and compelling products vs. Nvidia's equivalent offerings at the time.

As for the people questioning the source. MyDrivers broke the news about low-end Geforce GTX 1030 a week ago. Now PConline confirmed the card does exist. The same website leaked info about AMD's plans to refresh their mainstream cards with revamped Polaris GPUs (also confirming Polaris 12 at the occasion) and even got the launch date right (April 18) more than a month ago - all the usual websites quoted them (1, 2).

Is there any mention of an existing GTX 1040???

 

[Sweepr]

Is there any mention of an existing GTX 1040???

GeForce GT/GTX 1030 looks like the only GP108 card for desktops based on the websites above and AIDA64 release notes.

 

[tviceman]

GeForce GT/GTX 1030 looks like the only GP108 card for desktops based on the websites above and AIDA64 release notes.

What I mean is, is there an existence of a gt/gtx 1040?

 

[Bacon1]

Poor AMD; they should have reverted back to HD 300/400/500 series cards and stuck with making the best possible mid-range dies possible.

Isn't that Polaris? Which everyone here hates them because they only did mid-range and not full top-bottom stack?

 

[itsmydamnation]

AMD is planning to move to a new generation (Polaris -> Vega) in about 13 months. If AMD can do it, I don't see why NVIDIA can't pull it off, too.

Also, there was a lot of derision of Pascal as "Paxwell" because the underlying uArch was very similar to Maxwell's (with some enhancements & implemented in 16FF+), so it's not too farfetched to expect NVIDIA to move to a "true" new architecture 15 months-ish after Pascal.

But thats because 28nm maxwell should have never existed but 20nm was so crap.

if you consider that:
NV backported ( because they could afford it)
AMD forwardported (because they couldn't afford to and develop a in 2016 14nm part)

then everything aligns.
Volta will be what Volta is but it has nothing to do with maxwell and pascal being so similar, thats 20nm TSMC's fault.

 

[JDG1980]

I'm skeptical of this news. Nvidia has been on a steady 2-year cadence from Kepler onward.

• The first consumer Kepler GPU (GTX 680) was released in March 2012.
• The first consumer Maxwell GPU (GTX 750 Ti) was released in February 2014.
• The first consumer Pascal GPU (GTX 1080) was released in May 2016.

Could Nvidia release Volta early? Possibly, if it's ready. But if they do that, then what next? They are at the mercy of TSMC for node shrinks. As for architectural changes, I doubt they can consistently pull off the kind of 30%-50% performance gains on the same node that they got with Maxwell - implementing tiled rendering is a one-time thing. Why would Nvidia break a cadence that's working well for them, and risk having a longer than usual future period with nothing to release? I doubt very much that slow Q1 sales would be a reason. Q1 sales always suck, since they're coming down off the holidays.

If this is true, then I think it's for one of two reasons:

1. AMD's Vega is performing above expectations and beats GP102 by a substantial margin, and Nvidia wants to take back the performance crown as soon as possible.
2. Volta is actually a smaller jump than Kepler->Maxwell or Maxwell->Pascal. It could be something more along the lines of Skylake->Kaby Lake (to use a CPU-side analogy), with most of the gains coming from higher clock speeds due to the process tweaks. (Volta is said to be on TSMC "12nm", but that process is basically just a refined version of 16FF+.)

 

[JDG1980]

AMD is planning to move to a new generation (Polaris -> Vega) in about 13 months. If AMD can do it, I don't see why NVIDIA can't pull it off, too.

The difference is that AMD didn't release a full Polaris stack. Polaris was, in retrospect, basically a pipe cleaner for GloFo 14nm. AMD needs Vega as soon as possible because there's nothing to compete at the high end otherwise.

AMD seems to prefer incremental improvements with 1-2 new chips each year rather than Nvidia's approach of a full stack every 2 years. GCN 1.1 in 2013, GCN 1.2 in 2014, first HBM GPU in 2015, Polaris in 2016. None of these were full-stack releases, and Vega apparently won't be either (Polaris Refresh will still hold down the low-mid range).

Assuming we're looking at 16nm/12nm I'm guessing Volta GV104 will be to GP102 that Maxwell GM204 was to GK110 - about 10% faster, which will still make it impossible for AMD to compete with Vega being way bigger in die size, way more power hungry, and strapped with way more expensive HBM2 memory. Poor AMD; they should have reverted back to HD 300/400/500 series cards and stuck with making the best possible mid-range dies possible. The 4850/4870 and 5850/5870 were very robust and compelling products vs. Nvidia's equivalent offerings at the time.

Let's see what AMD has to offer with Vega before counting them out.

I'm also a bit less confident than you that Nvidia can nonchalantly pull off another 30%-50% increase in performance per die area without a full node shrink. Just because they did it once with Maxwell doesn't mean they can do it consistently every time. If Volta is a truly new ground-up design (which would be Nvidia's first since Fermi) then I suppose it's possible, but even then it would be tough.

 

[Ajay]

Volta is actually a smaller jump than Kepler->Maxwell or Maxwell->Pascal. It could be something more along the lines of Skylake->Kaby Lake (to use a CPU-side analogy), with most of the gains coming from higher clock speeds due to the process tweaks. (Volta is said to be on TSMC "12nm", but that process is basically just a refined version of 16FF+.)

I had seen some info (which I can't find ATM) that TSMC's "12nm" was ~15% more dense than 16FF+. I can only guess that this process node is has better performance as well (or lower power @ iso speed) that would make it worth while for GV100 that is going into ORNL's Summit in early 2018 (along with Power9). I'll keep grinding on google till I can find more info on the updated 16FF+ node.

 

[xpea]

...GV100 that is going into ORNL's Summit in early 2018 (along with Power9).

I don't know why people keep repeating this false information. Nvidia will start shipping Volta this summer into ORNL, not in 2018.
2018 is the official opening to users of the new supercomputer. But this kind of machine needs a very long qualification time before going online.
HPC GV100 Volta has been taped out long time ago and it must go into full production soon to meed ORNL calendar.
Now the important question for us, gamers, is what's GV104 schedule. Well it all depends on Nvidia priority. Do they want laptops first to maximize revenue or desktop first to maintain their performance leadership, thus their high margins ?
At this stage, I have no idea...

 

[tviceman]

Isn't that Polaris? Which everyone here hates them because they only did mid-range and not full top-bottom stack?

What's special about Polaris when compared to past AMD chips from the era I mentioned? It's bigger, consumes more power, and requires more resources then it's competition without being faster. Polaris's price from the start wasn't compelling either as it slotted in to be an average value option. The HD 4000/5000 series chips were all leaner and more efficient than their direct competition and were price winners out of the gate.

 

[tviceman]

Let's see what AMD has to offer with Vega before counting them out.

Nvidia went entirely unchallenged at the high end for a year. IF Vega manages to offer better performance or at least enticingly better perf/$, they'll have a much smaller window of opportunity to sell before getting stomped all over (again) by Volta. And even during Vega's "window of opportunity" Vega won't be alone in the high end space unlike the GTX 1070/1080/1080 TI /Titan X / Titan Xp that had zero competition.

I'm also a bit less confident than you that Nvidia can nonchalantly pull off another 30%-50% increase in performance per die area without a full node shrink.

Actually Maxwell day 1 reviews showed that GM204 had 25% better performance per mm2, not 30-50% like you're saying. If Nvidia can achieve 15-20% more performance per mm2 with Volta ( reasonable given architecture and node improvements) and they increase die size by 20%, that's about 10-15% faster than 1080 TI on a 378mm2 die.

Seems doable to me.

 

[Qwertilot]

Yup, its very plausible. The big thing they have to do to make it work is to keep the power draw under control while growing the chip sizes a bit.

 

[Head1985]

I doubt volta will be 12nm.TSMC is not ready for 300mm2+++ SKU.
GV104 will probably launch q1-q2 2018 at same 16FF node like pascal.

 

[Ajay]

I don't know why people keep repeating this false information. Nvidia will start shipping Volta this summer into ORNL, not in 2018.
2018 is the official opening to users of the new supercomputer. But this kind of machine needs a very long qualification time before going online.
HPC GV100 Volta has been taped out long time ago and it must go into full production soon to meed ORNL calendar.
Now the important question for us, gamers, is what's GV104 schedule. Well it all depends on Nvidia priority. Do they want laptops first to maximize revenue or desktop first to maintain their performance leadership, thus their high margins ?
At this stage, I have no idea...

If you have more information on GV100 shipping this summer - please share! We know that super computers have a long qualification period.

I know Ashraf Eassa thought GV100 was taping out in the summer of 2016, but I haven't found any confirmation Early indication are that GV100 is implemented on TSMC's "12nm" (enhanced 16FF+) node. This is important to gamers, because it means NV will already be receiving Volta GPU deliveries in 2017 . Which just lends credence to the idea that Volta can arrive sooner than than expected. If TSMC can produce a large die Volta on the new process, then GV104 is not likely far behind.

 

[xpea]

If you have more information on GV100 shipping this summer - please share! We know that super computers have a long qualification period.

Well I'm the one who leaked GV100 on TSMC 12nm node, information that every single tech website talked about and that was confirmed by TSMC investor meetings few weeks after. So yes I have my sources, but no I won't tell more

 

[Ajay]

Awww....

 

[sm625]

According to chinese website MyDrivers:

- Pascal lineup is complete now that NVIDIA launched Titan Xp and is about to launch GeForce GTX 1030

Is the GTX 1030 real? I thought it was just an april fools joke:

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

 

[railven]

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

W2 Edition would add a good ring to it.