NVIDIA Volta Rumor Thread

[TheF34RChannel]

It should have incentive, the fact that gta v can't sustain 4k 60 on max settings on the 1080 Ti
It's a 2015 game..

That's just one game (although there are surely more I cannot recall). Fact remains GDDR6 won't be ready until next March. Additionally, 4K is a very small market and currently basically requiring SLI Ti/Titan. 1440p is the sweetspot between the two resolutions for the moment for high end GPUs.

 

[nvgpu]

http://www.anandtech.com/show/11398...am-gddr6-added-to-catalogue-gddr5-gets-faster

SK Hynix has added GDDR6 memory chips to its product catalogue, revealing their general specifications and launch timeframe sometimes in Q4 2017.

GDDR6 is available by Q4 this year.

 

[Qwertilot]

Their incentive is the need to keep the upgrade cycle running of course.

Also, though, they will presumably want to keep the HPC and gaming stuff vaguely in sync, and they're having to move fast with the HPC stuff as there's some really fierce competition there. If they wait until Spring 2018 to launch any gaming Volta, then it'll only have something like a year before the next gen HPC part arrives, probably really denting how easily they can sell those gaming chips when it does so.

Anyway, we'll see Maybe they'll manage to keep the cycles quite distinct.

 

[Samwell]

Also, though, they will presumably want to keep the HPC and gaming stuff vaguely in sync, and they're having to move fast with the HPC stuff as there's some really fierce competition there. If they wait until Spring 2018 to launch any gaming Volta, then it'll only have something like a year before the next gen HPC part arrives, probably really denting how easily they can sell those gaming chips when it does so.

It's not just HPC stuff, the Volta successor gaming chips will launch in Mid 2019 in 7nm. So every Volta gaming gpu should be out in ~Mid 2018 to have enough time to make money. 2080Ti this time will come ~3months after the titan as it was with Maxwell. As for GV104, maybe it'll even launch this year for christmas time as Hynix seems to have GDDR6 available in Q4.

 

[crisium]

Citation for that that launch window for the still unnamed Volta successor?

 

[TheF34RChannel]

Their incentive is the need to keep the upgrade cycle running of course.

Also, though, they will presumably want to keep the HPC and gaming stuff vaguely in sync, and they're having to move fast with the HPC stuff as there's some really fierce competition there. If they wait until Spring 2018 to launch any gaming Volta, then it'll only have something like a year before the next gen HPC part arrives, probably really denting how easily they can sell those gaming chips when it does so.

Is it at this release rate? Nah...too soon I'm convinced.

Which HPC competition? Vega? Hardly worth mentioning that architecture. (Not trying to start a Volta vs Vega discussion here.) Additionally, I don't see what HPC releases have to do with consumer cards release schedules.

It's not just HPC stuff, the Volta successor gaming chips will launch in Mid 2019 in 7nm. So every Volta gaming gpu should be out in ~Mid 2018 to have enough time to make money. 2080Ti this time will come ~3months after the titan as it was with Maxwell. As for GV104, maybe it'll even launch this year for christmas time as Hynix seems to have GDDR6 available in Q4.

Are these your own thoughts or actual and factual information? I'm leaning towards the former. Even though Hynix may have GDDR6 available in Q4 2017 they do need to stockpile it, which won't be completed before Q1 2018.

 

[Pinstripe]

I thought they were doing a 12nm Pascal refresh with GDDR5X this year?

 

[Qwertilot]

Is it at this release rate? Nah...too soon I'm convinced.

Which HPC competition? Vega? Hardly worth mentioning that architecture. (Not trying to start a Volta vs Vega discussion here.) Additionally, I don't see what HPC releases have to do with consumer cards release schedules.

The HPC competition? Not AMD. Intel is the huge one I guess - they've huge resources to throw behind Knights landing & what comes after it. There's also dedicated machine learning stuff like Google at least have been playing with.

The HPC/Gaming release schedules still share a lot of the R&D work with the HPC stuff. Obviously now they've split off the actual chips, the HPC chips will arrive larger and sooner - vastly higher margins so they can afford a much less reliable process etc - but they won't want a huge gap.

I've no idea though. We'll see

 

[tviceman]

I thought they were doing a 12nm Pascal refresh with GDDR5X this year?

For all we know, Volta consumer cards are just that - Pascal refreshed with a more updated node. In the end, it'll be just like going from Kepler to Maxwell and Maxwell to Pascal - a 50-75% perf/w increase for the same tier of performance.

 

[TheF34RChannel]

I thought they were doing a 12nm Pascal refresh with GDDR5X this year?

For all we know, Volta consumer cards are just that - Pascal refreshed with a more updated node. In the end, it'll be just like going from Kepler to Maxwell and Maxwell to Pascal - a 50-75% perf/w increase for the same tier of performance.

Don't be silly both @Pinstripe there are currently 5 pages here alone saying otherwise, thus you may had noticed this is not the case before asking this now long forgotten silly 'article' that mused about that. Reading (most of) the thread is a worthwhile endeavor I highly recommend.

 

[tviceman]

Don't be silly both @Pinstripe there are currently 5 pages here alone saying otherwise, thus you may had noticed this is not the case before asking this now long forgotten silly 'article' that mused about that. Reading (most of) the thread is a worthwhile endeavor I highly recommend.

The point I was trying to make is that it doesn't matter whatever coming next is called. Nvidia's march has been very consistent since Kepler. Maxwell was 50-60% faster than Kepler at release, and Pascal was 50-60% faster than Maxwell. Expect the same out of whatever is next. The name doesn't matter.

 

[jpiniero]

The point I was trying to make is that it doesn't matter whatever coming next is called. Nvidia's march has been very consistent since Kepler. Maxwell was 50-60% faster than Kepler at release, and Pascal was 50-60% faster than Maxwell. Expect the same out of whatever is next. The name doesn't matter.

That's going to be tough without a shrink, and you don't really get all that much from 12 nm in terms of density. I think they've run out of large clock speed gains too, when you factor that most 1080 Tis can boost into at least the 1800 range.

I'm of the opinion that it will be a minor Pascal update on 12 nm; maybe an extra SM or two per die and slightly higher clocks. Consumer Volta will be later.

 

[Bouowmx]

maybe an extra SM or two per die and slightly higher clocks.

I don't know: that sounds pointless to me. Designing 3 new dies for 256 extra cores? Might as well design dies for Volta, featuring 140% the number of Pascal cores, as shown in GV100.

 

[Qwertilot]

That's surely(?) essentially what they're going to do. Needs a huge perf/watt uplift to work of course, but from V100 & Xavier they seem to have that with Volta.

Should give the 'expected' 50-60% so why wouldn't they?

 

[Samwell]

Citation for that that launch window for the still unnamed Volta successor?

No Citation, by looking at past release shedules and what AMD does.

Are these your own thoughts or actual and factual information? I'm leaning towards the former. Even though Hynix may have GDDR6 available in Q4 2017 they do need to stockpile it, which won't be completed before Q1 2018.

Thoughts, but it's the most logical in my view. In 28nm last maxwell chip , 980Ti came in Mid 2015. At that time 16nm was starting but apple and mobile makers took first year of 16nm orders and then 1 year later the successor to the 980TI came. With 7nm it's the same and i expect same shedules. 7nm starting in 2018, mobile makers take all at tsmc. I year later Gx104 which will suceed GV102. AMD is aggressively pushing 7nm as posted by their roadmaps. They'll have 7m gpus out in H1 2019 definitely, probably even in early 2019 and you can be sure that nvidia won't just look at them and let them take market share and make money, because of the node advantage. I could imagine Nvidia won't make a full chip lineup with volta, as GM107/GM108 came a bit later, but as i said i don't expect any more of the bigger voltas launching after mid 2018.

Like Micron needed to stockpile GDDR5X with 1080 launch, where no one believed it could be so early? If they write it on their official roadmaps there are good chances, that it'll be possible to launch a product in Q4. Whether they really do it, i don't know, but i wouldn't call it impossible if the chip is ready. Maybe they even launch with GDDR5X, as Micron also has 14Gbps in Q4. Impossible to predict.

 

[TheF34RChannel]

A factor may be that, AFAIK, only Nvidia took GDDR5X, whereas GDDR6 will see more customers. Indeed difficult to predict and I remain skeptical until we are given the cold, hard facts.

 

[tviceman]

That's going to be tough without a shrink, and you don't really get all that much from 12 nm in terms of density. I think they've run out of large clock speed gains too, when you factor that most 1080 Tis can boost into at least the 1800 range.

I'm of the opinion that it will be a minor Pascal update on 12 nm; maybe an extra SM or two per die and slightly higher clocks. Consumer Volta will be later.

12nm doesn't offer much density improvement, but it offers substantial perf/w improvement (33%, I believe). Maxwell GM204 was 25% faster per mm2 than GK104, and that was on the exact same node. GV104 can be 360-400mm2 and bring performance better than GP102 with ~170w TDP.

Plenty of room for gains.

 

[TheF34RChannel]

Bit off topic, but why is every Ti so expensive? I never really looked at prices here and did just now for fun and: 965 EUR = 844.873 GBP as an average price...it's like the out of control RAM prices it seems? At this rate it's going to be Volta (regular and Ti) for Titan prices (not liking this potential foresight at all).

 

[Bouowmx]

NVIDIA GeForce GTX 1080 Ti at a normal 700 USD at Newegg: https://www.newegg.com/Product/ProductList.aspx?Submit=ENE&N=100007709 601294835 601295933&IsNodeId=1&bop=And&Order=PRICE&PageSize=96

Hopefully Volta pricing is no greater than Pascal's; Volta is looking like to use similar manufacturing process as Pascal's (should be mature now), but with improved power characteristics, and increased die area.

 

[TheF34RChannel]

NVIDIA GeForce GTX 1080 Ti at a normal 700 USD at Newegg: https://www.newegg.com/Product/ProductList.aspx?Submit=ENE&N=100007709 601294835 601295933&IsNodeId=1&bop=And&Order=PRICE&PageSize=96

Hopefully Volta pricing is no greater than Pascal's; Volta is looking like to use similar manufacturing process as Pascal's (should be mature now), but with improved power characteristics, and increased die area.

I don't live in the US so Newegg means nothing to me. It's good to see that it's not overpriced everywhere though A guy from the UK got his for about 660 GBP and now, a mere 2 weeks later it's around 840 GBP = insanity!

They need to come down before we get consumer Volta, I think, because there is no chance they'll price it lower than the previous generation. Might be the mining craze we need to get rid off.

 

[zuzu]

well by nvidia volta is out for like 4 years now )
i rly hope they will plop out dx12/vulcan hw spec in this gpu ,
i mean they say volta will be 2014 gpu (dx11)
so this is 2018 gpu and all game ind moving to hw vulcan dx12
if it plop out like just another dx11 gpu like pascal ,,
huh i will start to slowly hate nvidia

i just wish rx vega to kill all gtx gpus till now just so we can get best gpu ever in volta

 

[Head1985]

Bit off topic, but why is every Ti so expensive? I never really looked at prices here and did just now for fun and: 965 EUR = 844.873 GBP as an average price...it's like the out of control RAM prices it seems? At this rate it's going to be Volta (regular and Ti) for Titan prices (not liking this potential foresight at all).

mining..god i hate miners..
Everything from 1050TI to 1080TI is sold out or cost insane money.Maybe not in US, but in EU and every other country.

 

[tviceman]

Bit off topic, but why is every Ti so expensive? I never really looked at prices here and did just now for fun and: 965 EUR = 844.873 GBP as an average price...it's like the out of control RAM prices it seems? At this rate it's going to be Volta (regular and Ti) for Titan prices (not liking this potential foresight at all).

I'm guessing we'll see a $599-649 USD price tag for GTX 2080. If it's only ~10% faster than the GTX 1080 TI, then I think $599 is more likely. If GV204 is able to crush 1080 TI by 20% or more at 4k, then expect $649 or more pricing.

 

[jpiniero]

There was a rumor that the top GV104 would be sold as a Titan first, and priced accordingly.

 

[Arachnotronic]

There was a rumor that the top GV104 would be sold as a Titan first, and priced accordingly.

Doubtful.