[TT]Volta-based GeForce cards coming in 2018 with 48GB GDDR6

[Jaydip]

http://www.tweaktown.com/news/54148/volta-based-geforce-cards-coming-2018-48gb-gddr6/index.html

Moar ram moar

 

[Bacon1]

Title: Volta-based GeForce cards coming in 2018 with 48GB GDDR6

Article: The new 16GB GDDR6 cards will be based on the upcoming Volta-based graphics cards with a 256-bit memory bus, while the higher-end GV102 will use a faster 384-bit memory bus and possibly 24GB or even 48GB cards thanks to GDDR6.

This is why rumor sites drive me crazy.

"Who knows, might even use 64GB of GDDR6!!" Clickbait at its finest.

Edit: And the whole thing is based off a rumor that they will be announcing the cards May 2017 based off wccftech.

 

[mohit9206]

I'll bet $100 that there won't be any 48GB consumer gaming cards in 2018.I'll bet $50 there won't be any in 2019 either.And $20 there won't be any in 2020 either.Pure clickbait.
Best case scenario,it will be 12GB for flagships, 8GB for mid-range and 4GB for entry level.

 

[Qwertilot]

Well, we just got (the keynote on the main site) a promised 'late 2017' date for a 20w Volta based SoC. Suppose it does make you wonder if they might not slip some of the consumer cards out in 2017 sometime. Actually, I don't think they released the Pascal based SoC's until the main cards were done, so you could conclude they'll get them out next Spring.

Spring 2018 does seem more likely given they've still got big Pascal to really release, even GDDR5X isn't big volume yet etc.

 

[happy medium]

Kepler gtx680 = gtx1080
Kepler gtx780 = Pascal refresh coming in about 8 months.
maxwell gtx980 = Volta a year after the 1080 refresh with great power reduction ,higher clocks. 7nm or finfet+?

this release schedule seems very familiar. I don't think they will drag out the 14nm node as much as they did with 28nm.

 

[jpiniero]

Volta is TSMC 16FF

 

[happy medium]

Volta is TSMC 16FF

I think its just called finfet+ or something like that. I made correction thanks.

Vram? That's easy. Low end volta will have 6/8gb, 1080 replacement will have 16gb,and the Titan volta will have 24gb.

 

[jpiniero]

And they are going to drag it out. 2 years of Pascal, 2 years of Volta.

 

[happy medium]

And they are going to drag it out. 2 years of Pascal, 2 years of Volta.

Unless something awful happens, the Volta cards should be only 1 year with finfet+, and refreshed with 7nm.

 

[Qwertilot]

Are they going to have 7nm working well enough for big(ish!) GPU's by 2019? Seems an ambitious target with 10nm to get through first.

10nm they presumably will have, but that might not be enough of a jump for them to bother with it, as per 20nm.

 

[happy medium]

Are they going to have 7nm working well enough for big(ish!) GPU's by 2019? Seems an ambitious target with 10nm to get through first.

10nm they presumably will have, but that might not be enough of a jump for them to bother with it, as per 20nm.

they are skipping 10nm

 

[Sweepr]

All of this is based off info posted by Ishimura, known for previous NVIDIA leaks. I found out and posted here last thursday, then WCCFTech wrote a clickbait article about it yesterday - and now other websites are quoting WCCFTech instead of the original source(s).

 

[Sweepr]

Here's a teaser for Volta, the first product was announced today at GTC - next generation Xavier SoC (sampling Q4-17 at 16nm FF):

NVIDIA is promising 20 Deep Learning Tera-Ops (DL TOPS) at 20W.
- Same as Pascal-based Tegra PX 2 at 1/4 the power (their current 16nm FF product)
- 43% of what NVIDIA’s flagship Tesla P40 (GP102) can offer in a 250W card at 8% the power

 

[Ajay]

they are skipping 10nm

Who is skipping 10nm - NV? AFAIK, TSMC is doing a 10nm node.

 

[Qwertilot]

It'd be NV. Like they did with 20nm - these new designs are getting really expensive to do so it takes quite a big performance jump from the node to make them die shrink.

 

[railven]

http://www.tweaktown.com/news/54148/volta-based-geforce-cards-coming-2018-48gb-gddr6/index.html

Moar ram moar

Who needs system RAM at this point. Just run everything off the GPU!!!! Haha.

 

[Ajay]

It'd be NV. Like they did with 20nm - these new designs are getting really expensive to do so it takes quite a big performance jump from the node to make them die shrink.

Makes sense. Thanks!